CN110378438A - Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant - Google Patents
Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant Download PDFInfo
- Publication number
- CN110378438A CN110378438A CN201910727040.9A CN201910727040A CN110378438A CN 110378438 A CN110378438 A CN 110378438A CN 201910727040 A CN201910727040 A CN 201910727040A CN 110378438 A CN110378438 A CN 110378438A
- Authority
- CN
- China
- Prior art keywords
- pixel
- mass index
- training
- image
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
This disclosure relates to a kind of training method of the Image Segmentation Model under label is fault-tolerant, device and relevant device.This method comprises: obtaining training sample set；Sample image is handled by parted pattern, obtains prediction segmentation result；Segmentation loss function is determined according to the Pixel-level of prediction segmentation result and sample image mark；Sample image and its Pixel-level mark are handled by quality sensor model and anti-over-fitting model, obtain relative mass index；According to segmentation loss function and relative mass index, the parameter of parted pattern and quality sensor model is adjusted, obtains the parted pattern that training is completed.This disclosure relates to label it is fault-tolerant under Image Segmentation Model training method, device and relevant device, relative mass index is generated according to sample image and its Pixel-level mark, and segmentation loss function is adjusted to complete model training according to relative mass index, it can still guarantee the parted pattern accuracy rate with higher after training when training sample set has noise.
Description
Technical field
This disclosure relates to field of computer technology, in particular to a kind of Image Segmentation Model of label under fault-tolerant
Training method, device, electronic equipment and computer-readable medium.
Background technique
In the field of medical imaging, marks quality and depend heavily on professional knowledge abundant, internet great amount of images data
Crowdsourcing labeling form be not appropriate for, and different doctor's labeled accustomed is different, causes used mark between different samples
Standard is not unification, and noise label becomes generally existing problem.The mark quality uneven for level, if directly adopted
It is trained with original mark, obtained model will receive the influence of noise sample, lead to final prediction result precision
Low, model robustness and generalization ability are poor.The state of an illness is diagnosed in view of doctor does not in clinical practice often need accurately to mark,
And for medical image segmentation problem, algorithm needs to carry out a large amount of accurate Pixel-level marks, therefore noise marks this problem
It is more severe in image segmentation；And existing scheme is the noise mark for being directed to image classification problem at present, therefore such as
What obtains the higher Image Segmentation Model of accuracy rate by the sample training of the uneven Pixel-level mark of level, is at present still not yet
It solves the problems, such as.
Therefore, it is necessary to training method, device, electronic equipment and the meters of a kind of Image Segmentation Model of new label under fault-tolerant
Calculation machine readable medium.
Above- mentioned information are only used for reinforcing the understanding to the background of the disclosure, therefore it disclosed in the background technology part
It may include the information not constituted to the relevant technologies known to persons of ordinary skill in the art.
Summary of the invention
In view of this, the embodiment of the present disclosure provide a kind of label it is fault-tolerant under the training method of Image Segmentation Model, device,
Electronic equipment and computer-readable medium can still guarantee that the parted pattern after training has when training sample set has noise
Higher accuracy rate.
Other characteristics and advantages of the disclosure will be apparent from by the following detailed description, or partially by the disclosure
Practice and acquistion.
According to the one side of the disclosure, a kind of training method of the Image Segmentation Model under label is fault-tolerant, this method are proposed
It include: acquisition training sample set, the training sample set includes sample image and its Pixel-level mark；By parted pattern to institute
It states sample image to be handled, obtains prediction segmentation result；According to the Pixel-level mark of the prediction segmentation result and sample image
Note determines segmentation loss function；The sample image and its Pixel-level mark are handled by quality sensor model, obtained
Absolute mass index；The absolute mass index is handled by anti-over-fitting model, obtains relative mass index；According to
The segmentation loss function and the relative mass index, adjust the parameter of the parted pattern and the quality sensor model,
To obtain the parted pattern that training is completed.
In a kind of exemplary embodiment of the disclosure, the method also includes: obtain image to be processed；Pass through described point
It cuts model to handle the image to be processed, obtains the Pixel-level segmentation result of the image to be processed.
In a kind of exemplary embodiment of the disclosure, obtaining training sample set includes: to obtain sample image and its accurate
Pixel-level mark；The pixel expansion that the first predetermined pixel value is carried out to the accurate Pixel-level mark of the first predetermined quantity, is obtained
The Pixel-level annotation results that must be extended；The picture of the second predetermined pixel value is carried out to the accurate Pixel-level mark of the second predetermined quantity
Element is shunk, and obtains the Pixel-level annotation results of contraction.
In a kind of exemplary embodiment of the disclosure, according to the segmentation loss function and the relative mass index,
The parameter of the parted pattern and the quality sensor model is adjusted, includes: root to obtain the parted pattern that training is completed
Target loss function is determined according to the product of the segmentation loss function and the relative mass index；According to the target loss letter
Number updates the parameter of the parted pattern and the quality sensor model, to obtain the parted pattern that training is completed.
In a kind of exemplary embodiment of the disclosure, by quality sensor model to the sample image and its Pixel-level
Mark is handled, and obtaining absolute mass index includes: to splice the sample image and its Pixel-level mark, is generated and is spelled
Connect data；The splicing data are handled by quality sensor model, obtain the absolute mass index.
In a kind of exemplary embodiment of the disclosure, by anti-over-fitting model to the absolute mass index at
Reason, obtaining relative mass index includes: to be handled by hyperbolic tangent function the absolute mass index, obtains opposite matter
Figureofmerit.
In a kind of exemplary embodiment of the disclosure, the quality sensor model includes convolutional neural networks model.
According to the one side of the disclosure, a kind of training device of the Image Segmentation Model under label is fault-tolerant, the device are proposed
It include: data acquisition module, for obtaining training sample set, the training sample set includes sample image and its Pixel-level mark
Note；Image segmentation module obtains prediction segmentation result for handling by parted pattern the sample image；Segmentation
Loss function computing module, for determining segmentation loss letter according to the Pixel-level mark of the prediction segmentation result and sample image
Number；Absolute mass Index module, for being handled by quality sensor model the sample image and its Pixel-level mark,
Obtain absolute mass index；Relative mass Index module, for being carried out by anti-over-fitting model to the absolute mass index
Processing obtains relative mass index；Parameter updating module, for according to the segmentation loss function and the relative mass index
The parameter of the parted pattern and the quality sensor model is adjusted, to obtain the parted pattern that training is completed.
According to the one side of the disclosure, a kind of electronic equipment is proposed, which includes: one or more processors；
Storage device, for storing one or more programs；When one or more programs are executed by one or more processors, so that one
A or multiple processors realize method as described above.
According to the one side of the disclosure, it proposes a kind of computer-readable medium, is stored thereon with computer program, the program
Method as described above is realized when being executed by processor.
According to the training method of Image Segmentation Model of the label of some embodiments offers of the disclosure under fault-tolerant, device, electricity
Sub- equipment and computer-readable medium generate absolute mass index according to sample image and its Pixel-level mark, can assess sample
The potential quality of the Pixel-level mark of this image；Processing is carried out to absolute mass index according to anti-over-fitting model and obtains opposite matter
Figureofmerit can reduce the gap of absolute mass index between the Pixel-level mark of different quality to a certain extent, avoid part
Sample image and its Pixel-level mark are endowed excessively high absolute mass index；And referred to according to segmentation loss function and relative mass
Mark adjusts the parameter of parted pattern to complete model training, can still guarantee the study of model when training sample set has noise
Performance, and make the parted pattern accuracy rate still with higher after training.
It should be understood that the above general description and the following detailed description are merely exemplary, this can not be limited
It is open.
Detailed description of the invention
Its example embodiment is described in detail by referring to accompanying drawing, above and other target, feature and the advantage of the disclosure will
It becomes more fully apparent.Drawings discussed below is only some embodiments of the present disclosure, for the ordinary skill of this field
For personnel, without creative efforts, it is also possible to obtain other drawings based on these drawings.
Fig. 1 is the training method of the Image Segmentation Model under a kind of label shown according to an exemplary embodiment is fault-tolerant
Flow chart.
Fig. 2 is the training method of the Image Segmentation Model under a kind of label for showing according to another exemplary embodiment is fault-tolerant
Flow chart.
Fig. 3 shows the treatment process schematic diagram of step S110 shown in Fig. 1 in one embodiment.
Fig. 4 shows the treatment process schematic diagram of step S160 shown in Fig. 1 in one embodiment.
Fig. 5 shows the treatment process schematic diagram of step S140 shown in Fig. 1 in one embodiment.
Fig. 6 is the schematic diagram of training sample set shown according to an exemplary embodiment.
Fig. 7 is the training structure of the Image Segmentation Model under a kind of label for showing according to a further exemplary embodiment is fault-tolerant
Schematic diagram.
Fig. 8 is the training device of the Image Segmentation Model under a kind of label shown according to an exemplary embodiment is fault-tolerant
Block diagram.
Fig. 9 is the block diagram of a kind of electronic equipment shown according to an exemplary embodiment.
Figure 10 is that a kind of computer readable storage medium schematic diagram is shown according to an exemplary embodiment.
Specific embodiment
Example embodiment is described more fully with reference to the drawings.However, example embodiment can be real in a variety of forms
It applies, and is not understood as limited to embodiment set forth herein；On the contrary, thesing embodiments are provided so that the disclosure will be comprehensively and complete
It is whole, and the design of example embodiment is comprehensively communicated to those skilled in the art.Identical appended drawing reference indicates in figure
Same or similar part, thus repetition thereof will be omitted.
In addition, described feature, structure or characteristic can be incorporated in one or more implementations in any suitable manner
In example.In the following description, many details are provided to provide and fully understand to embodiment of the disclosure.However,
It will be appreciated by persons skilled in the art that can with technical solution of the disclosure without one or more in specific detail,
Or it can be using other methods, constituent element, device, step etc..In other cases, it is not shown in detail or describes known side
Method, device, realization or operation are to avoid fuzzy all aspects of this disclosure.
Block diagram shown in the drawings is only functional entity, not necessarily must be corresponding with physically separate entity.
I.e., it is possible to realize these functional entitys using software form, or realized in one or more hardware modules or integrated circuit
These functional entitys, or these functional entitys are realized in heterogeneous networks and/or processor device and/or microcontroller device.
Flow chart shown in the drawings is merely illustrative, it is not necessary to including all content and operation/step,
It is not required to execute by described sequence.For example, some operation/steps can also decompose, and some operation/steps can close
And or part merge, therefore the sequence actually executed is possible to change according to the actual situation.
It should be understood that although herein various assemblies may be described using term first, second, third, etc., these groups
Part should not be limited by these terms.These terms are to distinguish a component and another component.Therefore, first group be discussed herein below
Part can be described as the second component without departing from the teaching of disclosure concept.As used herein, term " and/or " include associated
All combinations for listing any of project and one or more.
It will be understood by those skilled in the art that attached drawing is the schematic diagram of example embodiment, module or process in attached drawing
Necessary to not necessarily implementing the disclosure, therefore it cannot be used for the protection scope of the limitation disclosure.
Fig. 1 is the training method of the Image Segmentation Model under a kind of label shown according to an exemplary embodiment is fault-tolerant
Flow chart.The training method of Image Segmentation Model under the label that the embodiment of the present disclosure provides is fault-tolerant can be by arbitrarily having calculating
The electronic equipment of processing capacity executes, such as user terminal and/or server are executed in the following embodiments with server
It is illustrated for the method, but it's not limited to that for the disclosure.Under the label that the embodiment of the present disclosure provides is fault-tolerant
The training method 10 of Image Segmentation Model may include step S110 to S160.
As shown in Figure 1, in step s 110, obtain training sample set, the training sample set include sample image and its
Pixel-level mark.
In the embodiment of the present disclosure, the sample quality that the training sample is concentrated is not restricted by, i.e., the described sample image
Noise may be present in Pixel-level mark.Fig. 6 is the schematic diagram of training sample set shown according to an exemplary embodiment.In image point
It cuts in the mark sample of problem, segmentation noise is generally derived from zonal Pixel-level extension or shrinks.As shown in fig. 6, first row
For sample image, second is classified as accurate Pixel-level mark, and third is classified as the Pixel-level mark for shrinking noise, and the 4th is classified as and has
Expand the Pixel-level mark of noise.The training sample concentrates each sample image to have rib cage, lung, heart three organs
Pixel-level marks, and the schematic diagram of rib lung and heart is only shown in Fig. 6.It should be understood that the segmentation of the technical solution of the disclosure is appointed
Business is not limited to medical image, and Fig. 6 is only the example of training sample set, and training sample set can also be face image set, gesture
The other kinds of images such as image set, automobile image collection.
In the step s 120, the sample image is handled by parted pattern, obtains prediction segmentation result.
In the embodiment of the present disclosure, the parted pattern is used to carry out image segmentation operations to the sample image.Described point
Cutting model may be, for example, convolutional neural networks model.Wherein, convolutional neural networks (Convolutional Neural
Networks, CNN) it is a kind of comprising convolutional calculation and with the feedforward neural network of depth structure, it is deep learning (deep
Learning one of representative algorithm).Convolutional neural networks copy visual perception (visual perception) mechanism structure of biology
It builds, can exercise supervision study and unsupervised learning, the sparsity that the convolution kernel parameter sharing in hidden layer is connected with interlayer
Convolutional neural networks are enabled to reveal feature with lesser calculation amount plaid matching, such as pixel and audio are learnt, have stabilization
Effect and to data it is not additional Feature Engineering requirement.But the parted pattern of the disclosure can also be other kinds of model
The technical solution of structure, the disclosure to this and is not particularly limited.
In step s 130, segmentation loss letter is determined according to the Pixel-level mark of the prediction segmentation result and sample image
Number.
Wherein, loss function be that chance event or its value in relation to stochastic variable are mapped as nonnegative real number to indicate
The function of " risk " or " loss " of the chance event.In the application, loss function is usually as learning criterion and optimization problem
It is associated, i.e., by minimizing loss function solution and assessment models.Such as it is used for model in statistics and machine learning
Parameter Estimation.Segmentation loss function in the embodiment of the present disclosure is the parted pattern to the sample image and its Pixel-level
The loss function that mark obtains after being handled.
In step S140, the sample image and its Pixel-level mark are handled by quality sensor model, obtained
Obtain absolute mass index.
In the embodiment of the present disclosure, the quality sensor model is used for the mark marked to the sample image and its Pixel-level
Quality is evaluated, and can be labeled in semantic level to sample image and its Pixel-level and be combined and handled, and is obtained absolute
Quality index.Wherein, the absolute mass index can characterize whether the sample image and its Pixel-level mark have noise,
And the severity of mark noise.
In one embodiment, the quality sensor model includes convolutional neural networks model.For example, quality sensor model
It can be visual geometric group network (Visual Geometry Group Network, VGG network), the visual geometric networking
Network is a kind of improved volume and neural network, it is to be understood that specific mould of the technical solution of the disclosure to quality sensor model
Type structure is simultaneously not particularly limited.
In the embodiment of the present disclosure, the input channel number of the quality sensor model can be n+m, and wherein n is total class of mark
Not Shuo, m be sample image channel number, n and m are positive integer.By taking Fig. 6 as an example, the classification of mark is respectively background, rib
Bone, heart, lung, totally four classes, sample image are the grayscale image in a channel.Therefore using the data set shown in Fig. 6 as training sample
When this collection, the input channel number of the quality sensor model is n+m=4+1=5.
In step S150, the absolute mass index is handled by anti-over-fitting model, obtains relative mass
Index.
In the embodiment of the present disclosure, the input of the anti-over-fitting model is absolute mass index, exports and refers to for relative mass
Mark.According to absolute mass index, training sample can be picked out and concentrate the higher Pixel-level mark of relative mass, but absolute mass
Index will likely be endowed too high or too low weighted value, for example, 0 in post-processing, this will make in the training process can not
Part sample is effectively learnt.The anti-over-fitting model of this step is by generating relative mass index and being used for subsequent place
Reason can still have the case where significant difference guaranteeing the quality index between the different sample image of quality and its Pixel-level mark
Under, avoid the occurrence of the situation that weighted value is too high or too low in subsequent training step.
In one embodiment, the absolute mass index can be handled by hyperbolic tangent function, is obtained opposite
Quality index.Wherein, hyperbolic tangent function is one kind of hyperbolic functions.Hyperbolic tangent function is generally write on mathematical linguistics
Tanh can also write a Chinese character in simplified form into th.
In one embodiment, relative mass index can be calculate by the following formula:
Γ(x；Y)=λ tanh (Θ (x；y)) (1)
Wherein, λ is hyper parameter, Θ (x；Y) (x is marked for sample image and its Pixel-level；Y) absolute mass index, Γ
(x；Y) (x is marked for sample image and its Pixel-level；Y) relative mass index.
In step S160, according to the segmentation loss function and the relative mass index, the parted pattern is adjusted
With the parameter of the quality sensor model, the parted pattern completed with acquisition training.
In the disclosed embodiment, operation can be weighted to the segmentation loss function according to the relative mass index,
So that the segmentation loss function after weighting operations is to the noisy sample image of tool and its Pixel-level mark with bigger
Response.
According to the training method of Image Segmentation Model of the label of disclosure embodiment offer under fault-tolerant, according to sample graph
Picture and its Pixel-level mark generate absolute mass index, can assess the potential quality of the Pixel-level mark of sample image；According to
Anti- over-fitting model carries out processing to absolute mass index and obtains relative mass index, can reduce not homogeneity to a certain extent
The gap of absolute mass index between the Pixel-level mark of amount avoids part sample image and its Pixel-level mark from being endowed excessively high
Absolute mass index；And the parameter of parted pattern is adjusted to complete model instruction according to segmentation loss function and relative mass index
Practice, can still guarantee the learning performance of model when training sample set has noise, and there is the parted pattern after training still
Higher accuracy rate.
It will be clearly understood that the present disclosure describes how to form and use particular example, but the principle of the disclosure is not limited to
These exemplary any details.On the contrary, the introduction based on disclosure disclosure, these principles can be applied to many other
Embodiment.
Fig. 2 is the training method of the Image Segmentation Model under a kind of label for showing according to another exemplary embodiment is fault-tolerant
Flow chart.
As shown in Fig. 2, the training method of the Image Segmentation Model under the label of the embodiment of the present invention is fault-tolerant can also be into one
Step includes the following steps.
In step S202, image to be processed is obtained.
In the embodiment of the present disclosure, the image to be processed can be the image for needing be split.For example, can be set by terminal
Standby or server obtains task to be split, includes one or more images to be processed in band segmentation task.
In step S204, the image to be processed is handled by the parted pattern, is obtained described to be processed
The Pixel-level segmentation result of image.
In the embodiment of the present disclosure, the parted pattern can be the parted pattern of training completion in Fig. 1.
In one embodiment, image can be handled by image, semantic segmentation evaluation index, for example, can pass through
Dice index is as evaluation criterion:
Wherein, VpredFor the cut zone of Pixel-level segmentation result, VgtFor tab area.Dice is evaluation index, Dice
Numerical value minimum value be 0, maximum value 1, numerical value is better closer to the segmentation effect of 1 representative model.
Fig. 3 shows the treatment process schematic diagram of step S110 shown in Fig. 1 in one embodiment.
As shown in figure 3, above-mentioned steps S110 may further include following steps in the embodiment of the present invention.
In step S111, sample image and its accurate Pixel-level mark are obtained.
In step S112, the pixel of the first predetermined pixel value is carried out to the accurate Pixel-level mark of the first predetermined quantity
Expansion, obtains the Pixel-level annotation results of extension.
In one embodiment, the first predetermined quantity may be less than the quantity for being equal to accurate Pixel-level mark it is any just
Integer.For example, the first predetermined quantity can be 0%, 25%, 50%, 75% etc. of the quantity of accurate Pixel-level mark, the disclosure
Technical solution to this and be not particularly limited.
In one embodiment, first predetermined pixel value may be less than the Pixel Dimensions equal to the sample image
Any positive integer.For example, the first predetermined pixel value can be 0-8,5-13 etc., the technical solution of the disclosure does not make special limit to this
It is fixed.
In step S113, the pixel of the second predetermined pixel value is carried out to the accurate Pixel-level mark of the second predetermined quantity
It shrinks, obtains the Pixel-level annotation results of contraction.
In the embodiment of the present disclosure, the second predetermined quantity, the second predetermined pixel value value rule respectively with the first predetermined number
Amount, the first predetermined pixel value are identical, and details are not described herein again.
The training method of Image Segmentation Model in the case where the label of the embodiment of the present disclosure is fault-tolerant, by the first predetermined quantity
Accurate Pixel-level mark carry out the pixel expansion of the first predetermined pixel value, and to the accurate picture of the second predetermined quantity
Plain grade mark carries out the pixel shrinkage of the second predetermined pixel value, can obtain the training with different noise types and level of noise
Sample set, and then parted pattern can be trained according to the training sample set, obtain the parted pattern with error resilience performance.
Fig. 4 shows the treatment process schematic diagram of step S160 shown in Fig. 1 in one embodiment.
As shown in figure 4, above-mentioned steps S160 may further include following steps in the embodiment of the present invention.
In step S161, target loss is determined according to the product of the segmentation loss function and the relative mass index
Function.
In the embodiment of the present disclosure, the target loss function can be determined by following formula:
Loss=∑I=1 ..., NΘ(xi, yi)·Li (3)
Wherein, LiTo divide loss function, Θ (xi, yi) it is relative mass index, i is that the sample that training sample is concentrated is compiled
Number, and meet s.t. ∑I=1 ..., NΘ(xi, yi)=1,0≤Θ (xi, yi)≤1。
In step S162, the parted pattern and the quality sensor model are updated according to the target loss function
Parameter, to obtain the parted pattern that training is completed.
In the embodiment of the present disclosure, the target loss function can be subjected to gradient passback, while adjusting the parted pattern
With the parameter of quality sensor model.
Fig. 5 shows the treatment process schematic diagram of step S140 shown in Fig. 1 in one embodiment.
As shown in figure 5, above-mentioned steps S140 may further include following steps in the embodiment of the present invention.
In step s 141, the sample image and its Pixel-level mark are spliced, generates splicing data.
The sample image can be the grayscale image with a channel in one embodiment, can also be the coloured silk with triple channel
Chromatic graph piece.Processing can be carried out to sample image and its Pixel-level mark splicing according to channel form and generate splicing data.
In step S142, the splicing data are handled by quality sensor model, obtain the absolute mass
Index.
The training method of Image Segmentation Model under the label that the embodiment of the present disclosure provides is fault-tolerant, not to training sample
Collection be trained in the case where further marking to model, alleviates noise label to the situation of segmentation task；And it can
It is trained according to the sample training collection with different type and different degrees of noise mark, obtains robustness and generalization ability
Higher parted pattern.The training method of Image Segmentation Model under the label that the embodiment of the present disclosure provides is fault-tolerant is able to solve mark
In the medical image field for infusing heavy dependence expertise, noise label training effect caused by model training is bad to be lacked
It falls into.
Fig. 7 is the training structure of the Image Segmentation Model under a kind of label for showing according to a further exemplary embodiment is fault-tolerant
Schematic diagram.
As shown in fig. 7, the model framework of the embodiment of the present disclosure include three parts: parted pattern, quality sensor model and
Anti- over-fitting model.
Wherein, parted pattern can be the convolutional neural networks model for generating segmentation result.Quality sensor model can
To be convolutional neural networks structural model, for the splicing of image and its Pixel-level mark (to be marked as input in Fig. 7
Classification 1, classification 2 ... classification n), and run parallel with parted pattern.For the weighted sample again in same batch, quality
Sensor model and anti-over-fitting model generate an absolute mass for sample image and its Pixel-level mark of each input and refer to
Mark is followed by softmax layers, generates relative mass index (Relative Score).Relative mass index and by segmentation network
The segmentation loss function of generation, which is multiplied, generates final target loss function.Back-propagating target loss function comes while adjusting point
Cut model and quality sensor model.
Quality sensor model can supervise the fall off rate of target loss function.After Weighted Loss Function again, divide mould
Block backpropagation and more focuses on the sample with higher weight based on the loss weighted again, while quality perceives
Module is based on identical loss value, backpropagation.
Another xiIt is that training sample concentrates i-ththA sample image, yiIt is its Pixel-level mark.LiIt is with N number of sample
In batch data, calculated by segmentation module i-ththA segmentation loss function.The disclosure picture that sample image and training is used
Plain grade is stitched together, and a quality sensor model Θ () based on VGG network is inputted, to extract high dimensional information.It is shown with Fig. 6
The training sample set including gray level image for, the input channel number of the quality sensor model is n+1, and wherein n is mark
Total classification number (n=4, respectively represents background, lung, heart, rib cage in the present embodiment), ' adds 1 ' to be the list of representative sample image
Channel grayscale image.Wherein, pond layer can be averaged with single channel to replace the last layer based on VGG network, evaluation is in the batch
In each sample weight Θ (x；y).Then for distinguishing the visible public affairs of calculation of the target loss function of different label qualities
Formula (3).
Quality sensor model can pick out the higher mark sample of relative mass and be trained, but in the training process
Quality assessment modules can be tended to assign high weight to individual accurate samples, so that whole network is in a small amount of data
Serious over-fitting.If some sample has been assigned too high or too low weight in training original state simultaneously, softmax layers
Gradient at this position has been similar to 0, so that training later period network can not effectively modify the weight of these samples, leads
Selection solidification is caused, final over-fitting degree is aggravated.
The anti-over-fitting module of the embodiment of the present disclosure can limit absolutely between quality sensor model and softmax layers
To the value of quality index, relative mass index is generated.The visible formula of the calculation of relative mass index (1).Anti- over-fitting
The quality index of output is narrowed down to (- λ, λ) (relative mass index) from (- ∞, ∞) range (absolute mass index) by model.?
After anti-over-fitting model and softmax layers of processing, in same a batch, the maximum ratio of two quality index is reduced from ∞
To e^2 λ.Guarantee will not fully be ignored any one sample by this numerical value, while is able to solve since the mistake of early stage is chosen
The too high or too low problem of weight in training process caused by selecting.
The training method of Image Segmentation Model under the label that the embodiment of the present disclosure provides is fault-tolerant, can mark in Pixel-level
In the case where of low quality, still guarantee that model carries out efficiently high-quality learning process.Wherein, quality sensor model can identify sample
The profound connection of inherent semanteme between this image and its Pixel-level mark, and then obtain the absolute of assessment Pixel-level mark quality
Quality index；Meanwhile processing is carried out to absolute mass index by anti-over-fitting model and obtains relative mass index, it can guarantee
The generalization ability of network avoids over-fitting, and then guarantees there are enough training samples.
It will be appreciated by those skilled in the art that realizing that all or part of the steps of above-described embodiment is implemented as being executed by CPU
Computer program.When the computer program is executed by CPU, above-mentioned function defined by the above method that the disclosure provides is executed
Energy.The program can store in a kind of computer readable storage medium, which can be read-only memory, magnetic
Disk or CD etc..
Further, it should be noted that above-mentioned attached drawing is only the place according to included by the method for disclosure exemplary embodiment
Reason schematically illustrates, rather than limits purpose.It can be readily appreciated that above-mentioned processing shown in the drawings is not indicated or is limited at these
The time sequencing of reason.In addition, be also easy to understand, these processing, which can be, for example either synchronously or asynchronously to be executed in multiple modules.
Following is embodiment of the present disclosure, can be used for executing embodiments of the present disclosure.It is real for disclosure device
Undisclosed details in example is applied, embodiments of the present disclosure is please referred to.
Fig. 8 is the training device of the Image Segmentation Model under a kind of label shown according to an exemplary embodiment is fault-tolerant
Block diagram.The training device 80 of Image Segmentation Model under the label that the embodiment of the present disclosure provides is fault-tolerant may include: data acquisition
Module 810, image segmentation module 820, segmentation loss function computing module 830, absolute mass Index module 840, relative mass
Index module 850 and parameter updating module 860.
In the training device 80 of the Image Segmentation Model under label is fault-tolerant, data acquisition module 810 can be used for obtaining instruction
Practice sample set, the training sample set includes sample image and its Pixel-level mark.
In the exemplary embodiment, the data acquisition module 810 may include accurate mark acquiring unit, extend to mark and obtain
It takes unit and shrinks mark acquiring unit.Wherein, accurately mark acquiring unit can be used for obtaining sample image and its accurate
Pixel-level mark；Extension mark acquiring unit can be used for carrying out first to the accurate Pixel-level mark of the first predetermined quantity predetermined
The pixel of pixel value is expanded, and the Pixel-level annotation results of extension are obtained；Shrinking mark acquiring unit can be used for the second predetermined number
The accurate Pixel-level mark of amount carries out the pixel shrinkage of the second predetermined pixel value, obtains the Pixel-level annotation results of contraction.
Image segmentation module 820 can be used for handling the sample image by parted pattern, obtain prediction segmentation
As a result.
Segmentation loss function computing module 830 can be used for the Pixel-level mark according to the prediction segmentation result and sample image
Note determines segmentation loss function.
Absolute mass Index module 840 can be used for marking the sample image and its Pixel-level by quality sensor model
It is handled, obtains absolute mass index.
In the exemplary embodiment, the quality sensor model can be convolutional neural networks model.
In the exemplary embodiment, absolute mass Index module 840 may include concatenation unit and absolute mass index list
Member.Wherein concatenation unit can be used for splicing the sample image and its Pixel-level mark, generate splicing data；Absolute matter
Figureofmerit unit can be used for handling the splicing data by quality sensor model, obtain the absolute mass index.
Relative mass Index module 850 can be used for handling the absolute mass index by anti-over-fitting model,
Obtain relative mass index.
In the exemplary embodiment, relative mass Index module 850 can be used for through hyperbolic tangent function to described absolute
Quality index is handled, and relative mass index is obtained.
Parameter updating module 860 can be used for adjusting described point according to the segmentation loss function and the relative mass index
The parameter of model and the quality sensor model is cut, to obtain the parted pattern that training is completed.
In the exemplary embodiment, parameter updating module 860 may include target loss function calculating unit and parameter more
New unit.Wherein, target loss function calculating unit can be used for according to the segmentation loss function and the relative mass index
Product determine target loss function；Parameter updating unit can be used for updating the parted pattern according to the target loss function
With the parameter of the quality sensor model, the parted pattern completed with acquisition training.
In one embodiment, the training device 80 of the Image Segmentation Model under label is fault-tolerant may also include acquisition image and obtain
Modulus block and dividing processing module.Wherein, image collection module can be used for image to be processed；Dividing processing module can be used for leading to
It crosses the parted pattern to handle the image to be processed, obtains the Pixel-level segmentation result of the image to be processed.
According to the training device of Image Segmentation Model of the label of disclosure embodiment offer under fault-tolerant, according to sample graph
Picture and its Pixel-level mark generate absolute mass index, can assess the potential quality of the Pixel-level mark of sample image；According to
Anti- over-fitting model carries out processing to absolute mass index and obtains relative mass index, can reduce not homogeneity to a certain extent
The gap of absolute mass index between the Pixel-level mark of amount avoids part sample image and its Pixel-level mark from being endowed excessively high
Absolute mass index；And the parameter of parted pattern is adjusted to complete model instruction according to segmentation loss function and relative mass index
Practice, can still guarantee the learning performance of model when training sample set has noise, and there is the parted pattern after training still
Higher accuracy rate.
Fig. 9 is the block diagram of a kind of electronic equipment shown according to an exemplary embodiment.
The electronic equipment 200 of this embodiment according to the disclosure is described referring to Fig. 9.The electronics that Fig. 9 is shown
Equipment 200 is only an example, should not function to the embodiment of the present disclosure and use scope bring any restrictions.
As shown in figure 9, electronic equipment 200 is showed in the form of universal computing device.The component of electronic equipment 200 can wrap
It includes but is not limited to: at least one processing unit 210, at least one storage unit 220, (including the storage of the different system components of connection
Unit 220 and processing unit 210) bus 230, display unit 240 etc..
Wherein, the storage unit is stored with program code, and said program code can be held by the processing unit 210
Row, so that the processing unit 210 executes described in this specification above-mentioned electronic prescription circulation processing method part according to this
The step of disclosing various illustrative embodiments.For example, the processing unit 210 can be executed such as Fig. 1, Fig. 2, Fig. 3, Fig. 4, figure
Step shown in 5.
The storage unit 220 may include the readable medium of volatile memory cell form, such as random access memory
Unit (RAM) 2201 and/or cache memory unit 2202 can further include read-only memory unit (ROM) 2203.
The storage unit 220 can also include program/practical work with one group of (at least one) program module 2205
Tool 2204, such program module 2205 includes but is not limited to: operating system, one or more application program, other programs
It may include the realization of network environment in module and program data, each of these examples or certain combination.
Bus 230 can be to indicate one of a few class bus structures or a variety of, including storage unit bus or storage
Cell controller, peripheral bus, graphics acceleration port, processing unit use any bus structures in a variety of bus structures
Local bus.
Electronic equipment 200 can also be with one or more external equipments 300 (such as keyboard, sensing equipment, bluetooth equipment
Deng) communication, can also be enabled a user to one or more equipment interact with the electronic equipment 200 communicate, and/or with make
Any equipment (such as the router, modulation /demodulation that the electronic equipment 200 can be communicated with one or more of the other calculating equipment
Device etc.) communication.This communication can be carried out by input/output (I/O) interface 250.Also, electronic equipment 200 can be with
By network adapter 260 and one or more network (such as local area network (LAN), wide area network (WAN) and/or public network,
Such as internet) communication.Network adapter 260 can be communicated by bus 230 with other modules of electronic equipment 200.It should
Understand, although not shown in the drawings, other hardware and/or software module can be used in conjunction with electronic equipment 200, including but unlimited
In: microcode, device driver, redundant processing unit, external disk drive array, RAID system, tape drive and number
According to backup storage system etc..
Through the above description of the embodiments, those skilled in the art is it can be readily appreciated that example described herein is implemented
Mode can also be realized by software realization in such a way that software is in conjunction with necessary hardware.Therefore, according to the disclosure
The technical solution of embodiment can be embodied in the form of software products, which can store non-volatile at one
Property storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.) in or network on, including some instructions are so that a calculating
Equipment (can be personal computer, server or network equipment etc.) executes the above method according to disclosure embodiment.
Figure 10 schematically shows a kind of computer readable storage medium schematic diagram in disclosure exemplary embodiment.
Refering to what is shown in Fig. 10, describing the program product for realizing the above method according to embodiment of the present disclosure
400, can using portable compact disc read only memory (CD-ROM) and including program code, and can in terminal device,
Such as it is run on PC.However, the program product of the disclosure is without being limited thereto, in this document, readable storage medium storing program for executing can be with
To be any include or the tangible medium of storage program, the program can be commanded execution system, device or device use or
It is in connection.
Described program product can be using any combination of one or more readable mediums.Readable medium can be readable letter
Number medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can be but be not limited to electricity, magnetic, optical, electromagnetic, infrared ray or
System, device or the device of semiconductor, or any above combination.The more specific example of readable storage medium storing program for executing is (non exhaustive
List) include: electrical connection with one or more conducting wires, portable disc, hard disk, random access memory (RAM), read-only
Memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disc read only memory
(CD-ROM), light storage device, magnetic memory device or above-mentioned any appropriate combination.
The computer readable storage medium may include in a base band or the data as the propagation of carrier wave a part are believed
Number, wherein carrying readable program code.The data-signal of this propagation can take various forms, including but not limited to electromagnetism
Signal, optical signal or above-mentioned any appropriate combination.Readable storage medium storing program for executing can also be any other than readable storage medium storing program for executing
Readable medium, the readable medium can send, propagate or transmit for by instruction execution system, device or device use or
Person's program in connection.The program code for including on readable storage medium storing program for executing can transmit with any suitable medium, packet
Include but be not limited to wireless, wired, optical cable, RF etc. or above-mentioned any appropriate combination.
Can with any combination of one or more programming languages come write for execute the disclosure operation program
Code, described program design language include object oriented program language-Java, C++ etc., further include conventional
Procedural programming language-such as " C " language or similar programming language.Program code can be fully in user
It calculates and executes in equipment, partly executes on a user device, being executed as an independent software package, partially in user's calculating
Upper side point is executed on a remote computing or is executed in remote computing device or server completely.It is being related to far
Journey calculates in the situation of equipment, and remote computing device can pass through the network of any kind, including local area network (LAN) or wide area network
(WAN), it is connected to user calculating equipment, or, it may be connected to external computing device (such as utilize ISP
To be connected by internet).
Above-mentioned computer-readable medium carries one or more program, when said one or multiple programs are by one
When the equipment executes, so that the computer-readable medium implements function such as: obtaining training sample set, the training sample set packet
Include sample image and its Pixel-level mark；The sample image is handled by parted pattern, obtains prediction segmentation result；
Segmentation loss function is determined according to the Pixel-level mark of the prediction segmentation result and sample image；Pass through quality sensor model pair
The sample image and its Pixel-level mark are handled, and absolute mass index is obtained；By anti-over-fitting model to described exhausted
Quality index is handled, relative mass index is obtained；According to the segmentation loss function and the relative mass index, adjust
The parameter of the parted pattern and the quality sensor model is saved, to obtain the parted pattern that training is completed.
It will be appreciated by those skilled in the art that above-mentioned each module can be distributed in device according to the description of embodiment, it can also
Uniquely it is different from one or more devices of the present embodiment with carrying out corresponding change.The module and/or unit of above-described embodiment
It can be merged into a module and/or unit, multiple modules and/or unit can also be further split into.
By the description of above embodiment, those skilled in the art is it can be readily appreciated that example embodiment described herein
It can also be realized in such a way that software is in conjunction with necessary hardware by software realization.Therefore, implemented according to the disclosure
The technical solution of example can be embodied in the form of software products, which can store in a non-volatile memories
In medium (can be CD-ROM, USB flash disk, mobile hard disk etc.) or on network, including some instructions are so that a calculating equipment (can
To be personal computer, server, mobile terminal or network equipment etc.) it executes according to the method for the embodiment of the present disclosure.
It is particularly shown and described the exemplary embodiment of the disclosure above.It should be appreciated that the present disclosure is not limited to
Detailed construction, set-up mode or implementation method described herein；On the contrary, disclosure intention covers included in appended claims
Various modifications and equivalence setting in spirit and scope.
In addition, structure shown by this specification Figure of description, ratio, size etc., only to cooperate specification institute
Disclosure, for skilled in the art realises that be not limited to the enforceable qualifications of the disclosure with reading, therefore
Do not have technical essential meaning, the modification of any structure, the change of proportionate relationship or the adjustment of size are not influencing the disclosure
Under the technical effect and achieved purpose that can be generated, it should all still fall in technology contents disclosed in the disclosure and obtain and can cover
In the range of.Meanwhile cited such as "upper" in this specification, " first ", " second " and " one " term, be also only and be convenient for
Narration is illustrated, rather than to limit the enforceable range of the disclosure, relativeness is altered or modified, without substantive change
Under technology contents, when being also considered as the enforceable scope of the disclosure.
Claims (10)
1. a kind of training method of the Image Segmentation Model under label is fault-tolerant characterized by comprising
Training sample set is obtained, the training sample set includes sample image and its Pixel-level mark；
The sample image is handled by parted pattern, obtains prediction segmentation result；
Segmentation loss function is determined according to the Pixel-level mark of the prediction segmentation result and sample image；
The sample image and its Pixel-level mark are handled by quality sensor model, obtain absolute mass index；
The absolute mass index is handled by anti-over-fitting model, obtains relative mass index；
According to the segmentation loss function and the relative mass index, the parted pattern and the quality sensor model are adjusted
Parameter, with obtain training complete the parted pattern.
2. the method as described in claim 1, which is characterized in that further include:
Obtain image to be processed；
The image to be processed is handled by the parted pattern, the Pixel-level for obtaining the image to be processed divides knot
Fruit.
3. the method as described in claim 1, which is characterized in that obtaining training sample set includes:
Obtain sample image and its accurate Pixel-level mark；
The pixel expansion that the first predetermined pixel value is carried out to the accurate Pixel-level mark of the first predetermined quantity, obtains the picture of extension
Plain grade annotation results；
The pixel shrinkage that the second predetermined pixel value is carried out to the accurate Pixel-level mark of the second predetermined quantity, obtains the picture of contraction
Plain grade annotation results.
4. the method as described in claim 1, which is characterized in that referred to according to the segmentation loss function with the relative mass
Mark, adjusts the parameter of the parted pattern and the quality sensor model, includes: to obtain the parted pattern that training is completed
Target loss function is determined according to the product of the segmentation loss function and the relative mass index；
The parameter of the parted pattern and the quality sensor model is updated according to the target loss function, has been trained with obtaining
At the parted pattern.
5. the method as described in claim 1, which is characterized in that by quality sensor model to the sample image and its pixel
Grade mark is handled, and is obtained absolute mass index and is included:
The sample image and its Pixel-level mark are spliced, splicing data are generated；
The splicing data are handled by quality sensor model, obtain the absolute mass index.
6. the method as described in claim 1, which is characterized in that carried out by anti-over-fitting model to the absolute mass index
Processing, obtaining relative mass index includes:
The absolute mass index is handled by hyperbolic tangent function, obtains relative mass index.
7. the method as described in claim 1, which is characterized in that the quality sensor model includes convolutional neural networks model.
8. a kind of training device of the Image Segmentation Model under label is fault-tolerant characterized by comprising
Data acquisition module, for obtaining training sample set, the training sample set includes sample image and its Pixel-level mark；
Image segmentation module obtains prediction segmentation result for handling by parted pattern the sample image；
Divide loss function computing module, divides for being determined according to the Pixel-level mark of the prediction segmentation result and sample image
Cut loss function；
Absolute mass Index module, for by quality sensor model to the sample image and its Pixel-level mark at
Reason obtains absolute mass index；
Relative mass Index module obtains opposite for being handled by anti-over-fitting model the absolute mass index
Quality index；
Parameter updating module, for according to the segmentation loss function and the relative mass index adjusting parted pattern and
The parameter of the quality sensor model, to obtain the parted pattern that training is completed.
9. a kind of electronic equipment characterized by comprising
One or more processors；
Storage device, for storing one or more programs；
When one or more of programs are executed by one or more of processors, so that one or more of processors are real
The now method as described in any in claim 1-7.
10. a kind of computer-readable medium, is stored thereon with computer program, which is characterized in that described program is held by processor
The method as described in any in claim 1-7 is realized when row.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910727040.9A CN110378438A (en) | 2019-08-07 | 2019-08-07 | Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910727040.9A CN110378438A (en) | 2019-08-07 | 2019-08-07 | Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110378438A true CN110378438A (en) | 2019-10-25 |
Family
ID=68258467
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910727040.9A Pending CN110378438A (en) | 2019-08-07 | 2019-08-07 | Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110378438A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796673A (en) * | 2019-10-31 | 2020-02-14 | Oppo广东移动通信有限公司 | Image segmentation method and related product |
| CN110889816A (en) * | 2019-11-07 | 2020-03-17 | 北京量健智能科技有限公司 | Image segmentation method and device |
| CN111311608A (en) * | 2020-02-05 | 2020-06-19 | 方军 | Method, apparatus and computer-readable storage medium for assessing wounds |
| CN111402278A (en) * | 2020-02-21 | 2020-07-10 | 华为技术有限公司 | Segmentation model training method, image labeling method and related device |
| CN111507419A (en) * | 2020-04-22 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Training method and device of image classification model |
| CN111583199A (en) * | 2020-04-24 | 2020-08-25 | 上海联影智能医疗科技有限公司 | Sample image annotation method and device, computer equipment and storage medium |
| CN112767422A (en) * | 2021-02-01 | 2021-05-07 | 推想医疗科技股份有限公司 | Training method and device of image segmentation model, segmentation method and device, and equipment |
| CN113643311A (en) * | 2021-06-28 | 2021-11-12 | 清华大学 | Image segmentation method and device for boundary error robustness |
| WO2023226606A1 (en) * | 2022-05-23 | 2023-11-30 | 北京百度网讯科技有限公司 | Image segmentation sample generation method and apparatus, method and apparatus for pre-training image segmentation model, and device and medium |
| CN113947561B (en) * | 2020-06-30 | 2023-12-26 | 应用材料以色列公司 | Image segmentation of semiconductor samples |
| CN117610263A (en) * | 2023-11-23 | 2024-02-27 | 上海百琪迈科技(集团)有限公司 | 3D visual discharging software |
| CN117876384A (en) * | 2023-12-21 | 2024-04-12 | 珠海横琴圣澳云智科技有限公司 | Target object instance segmentation and model training method and related products |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107123123A (en) * | 2017-05-02 | 2017-09-01 | 电子科技大学 | Image segmentation quality evaluating method based on convolutional neural networks |
| CN108596184A (en) * | 2018-04-25 | 2018-09-28 | 清华大学深圳研究生院 | Training method, readable storage medium storing program for executing and the electronic equipment of image, semantic parted pattern |
| CN108877839A (en) * | 2018-08-02 | 2018-11-23 | 南京华苏科技有限公司 | The method and system of perceptual evaluation of speech quality based on voice semantics recognition technology |
| CN108985269A (en) * | 2018-08-16 | 2018-12-11 | 东南大学 | Converged network driving environment sensor model based on convolution sum cavity convolutional coding structure |
| CN109086768A (en) * | 2018-07-13 | 2018-12-25 | 南京邮电大学 | The semantic image dividing method of convolutional neural networks |
| CN109308692A (en) * | 2018-07-30 | 2019-02-05 | 西北大学 | Based on the OCT image quality evaluating method for improving Resnet and SVR mixed model |
| CN109344833A (en) * | 2018-09-04 | 2019-02-15 | 中国科学院深圳先进技术研究院 | Medical image cutting method, segmenting system and computer readable storage medium |
| CN109978893A (en) * | 2019-03-26 | 2019-07-05 | 腾讯科技（深圳）有限公司 | Training method, device, equipment and the storage medium of image, semantic segmentation network |
| CN110070510A (en) * | 2019-04-26 | 2019-07-30 | 东北大学 | A kind of CNN medical image denoising method for extracting feature based on VGG-19 |
-
2019
- 2019-08-07 CN CN201910727040.9A patent/CN110378438A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107123123A (en) * | 2017-05-02 | 2017-09-01 | 电子科技大学 | Image segmentation quality evaluating method based on convolutional neural networks |
| CN108596184A (en) * | 2018-04-25 | 2018-09-28 | 清华大学深圳研究生院 | Training method, readable storage medium storing program for executing and the electronic equipment of image, semantic parted pattern |
| CN109086768A (en) * | 2018-07-13 | 2018-12-25 | 南京邮电大学 | The semantic image dividing method of convolutional neural networks |
| CN109308692A (en) * | 2018-07-30 | 2019-02-05 | 西北大学 | Based on the OCT image quality evaluating method for improving Resnet and SVR mixed model |
| CN108877839A (en) * | 2018-08-02 | 2018-11-23 | 南京华苏科技有限公司 | The method and system of perceptual evaluation of speech quality based on voice semantics recognition technology |
| CN108985269A (en) * | 2018-08-16 | 2018-12-11 | 东南大学 | Converged network driving environment sensor model based on convolution sum cavity convolutional coding structure |
| CN109344833A (en) * | 2018-09-04 | 2019-02-15 | 中国科学院深圳先进技术研究院 | Medical image cutting method, segmenting system and computer readable storage medium |
| CN109978893A (en) * | 2019-03-26 | 2019-07-05 | 腾讯科技（深圳）有限公司 | Training method, device, equipment and the storage medium of image, semantic segmentation network |
| CN110070510A (en) * | 2019-04-26 | 2019-07-30 | 东北大学 | A kind of CNN medical image denoising method for extracting feature based on VGG-19 |
Non-Patent Citations (1)
| Title |
|---|
| HAIDONG ZHU, ET AL: "《Pick-and Learn：Automatic Quality Evalution for Noisy-Labeled Image Segmentation》", 《ARXIV PREPRINT:ARXIV:1907.11835V1》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796673A (en) * | 2019-10-31 | 2020-02-14 | Oppo广东移动通信有限公司 | Image segmentation method and related product |
| CN110796673B (en) * | 2019-10-31 | 2023-02-24 | Oppo广东移动通信有限公司 | Image segmentation method and related product |
| CN110889816B (en) * | 2019-11-07 | 2022-12-16 | 拜耳股份有限公司 | Image segmentation method and device |
| CN110889816A (en) * | 2019-11-07 | 2020-03-17 | 北京量健智能科技有限公司 | Image segmentation method and device |
| CN111311608A (en) * | 2020-02-05 | 2020-06-19 | 方军 | Method, apparatus and computer-readable storage medium for assessing wounds |
| CN111311608B (en) * | 2020-02-05 | 2023-05-09 | 方军 | Method, apparatus and computer readable storage medium for evaluating wounds |
| CN111402278A (en) * | 2020-02-21 | 2020-07-10 | 华为技术有限公司 | Segmentation model training method, image labeling method and related device |
| CN111402278B (en) * | 2020-02-21 | 2023-10-27 | 华为云计算技术有限公司 | Segmentation model training method, image labeling method and related devices |
| CN111507419A (en) * | 2020-04-22 | 2020-08-07 | 腾讯科技（深圳）有限公司 | Training method and device of image classification model |
| CN111583199A (en) * | 2020-04-24 | 2020-08-25 | 上海联影智能医疗科技有限公司 | Sample image annotation method and device, computer equipment and storage medium |
| CN111583199B (en) * | 2020-04-24 | 2023-05-26 | 上海联影智能医疗科技有限公司 | Sample image labeling method, device, computer equipment and storage medium |
| CN113947561B (en) * | 2020-06-30 | 2023-12-26 | 应用材料以色列公司 | Image segmentation of semiconductor samples |
| CN112767422B (en) * | 2021-02-01 | 2022-03-08 | 推想医疗科技股份有限公司 | Training method and device of image segmentation model, segmentation method and device, and equipment |
| CN112767422A (en) * | 2021-02-01 | 2021-05-07 | 推想医疗科技股份有限公司 | Training method and device of image segmentation model, segmentation method and device, and equipment |
| CN113643311A (en) * | 2021-06-28 | 2021-11-12 | 清华大学 | Image segmentation method and device for boundary error robustness |
| CN113643311B (en) * | 2021-06-28 | 2024-04-09 | 清华大学 | Image segmentation method and device with robust boundary errors |
| WO2023226606A1 (en) * | 2022-05-23 | 2023-11-30 | 北京百度网讯科技有限公司 | Image segmentation sample generation method and apparatus, method and apparatus for pre-training image segmentation model, and device and medium |
| CN117610263A (en) * | 2023-11-23 | 2024-02-27 | 上海百琪迈科技(集团)有限公司 | 3D visual discharging software |
| CN117876384A (en) * | 2023-12-21 | 2024-04-12 | 珠海横琴圣澳云智科技有限公司 | Target object instance segmentation and model training method and related products |
| CN117876384B (en) * | 2023-12-21 | 2024-08-20 | 珠海横琴圣澳云智科技有限公司 | Target object instance segmentation and model training method and related products |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378438A (en) | Training method, device and the relevant device of Image Segmentation Model under label is fault-tolerant | |
| Sarkar et al. | Hands-On Transfer Learning with Python: Implement advanced deep learning and neural network models using TensorFlow and Keras | |
| WO2022007823A1 (en) | Text data processing method and device | |
| AU2017230184B2 (en) | Structure learning in convolutional neural networks | |
| US9965717B2 (en) | Learning image representation by distilling from multi-task networks | |
| Rodrigues et al. | Sequence labeling with multiple annotators | |
| EP3467723A1 (en) | Machine learning based network model construction method and apparatus | |
| CN114398961A (en) | Visual question-answering method based on multi-mode depth feature fusion and model thereof | |
| CN110196908A (en) | Data classification method, device, computer installation and storage medium | |
| CN113344206A (en) | Knowledge distillation method, device and equipment integrating channel and relation feature learning | |
| WO2021073390A1 (en) | Data screening method and apparatus, device and computer-readable storage medium | |
| CN109919077A (en) | Gesture recognition method, device, medium and calculating equipment | |
| CN113722474A (en) | Text classification method, device, equipment and storage medium | |
| CN115050064A (en) | Face living body detection method, device, equipment and medium | |
| Norris | Machine Learning with the Raspberry Pi | |
| CN114330499A (en) | Method, device, equipment, storage medium and program product for training classification model | |
| EP4220555A1 (en) | Training method and apparatus for image segmentation model, image segmentation method and apparatus, and device | |
| Sikka | Elements of Deep Learning for Computer Vision: Explore Deep Neural Network Architectures, PyTorch, Object Detection Algorithms, and Computer Vision Applications for Python Coders (English Edition) | |
| CN116451081A (en) | Data drift detection method, device, terminal and storage medium | |
| US20230360364A1 (en) | Compositional Action Machine Learning Mechanisms | |
| CN111382619B (en) | Picture recommendation model generation method, picture recommendation method, device, equipment and medium | |
| CN110738261B (en) | Image classification and model training method and device, electronic equipment and storage medium | |
| CN111651626A (en) | Image classification method and device and readable storage medium | |
| CN116680401A (en) | Document processing method, document processing device, apparatus and storage medium | |
| CN114116959B (en) | Aspect-level emotion analysis method and device and terminal |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191025 |
|
| RJ01 | Rejection of invention patent application after publication |