CN118116061A - Image processing system based on personnel identification - Google Patents
Image processing system based on personnel identification Download PDFInfo
- Publication number
- CN118116061A CN118116061A CN202410533876.6A CN202410533876A CN118116061A CN 118116061 A CN118116061 A CN 118116061A CN 202410533876 A CN202410533876 A CN 202410533876A CN 118116061 A CN118116061 A CN 118116061A
- Authority
- CN
- China
- Prior art keywords
- feature extraction
- extraction network
- face detector
- loss function
- parameters
- 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.)
- Granted
Links
- 238000012545 processing Methods 0.000 title claims abstract description 56
- 238000000605 extraction Methods 0.000 claims abstract description 179
- 238000012360 testing method Methods 0.000 claims abstract description 44
- 230000000694 effects Effects 0.000 claims abstract description 34
- 238000011156 evaluation Methods 0.000 claims abstract description 28
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 27
- 230000006870 function Effects 0.000 claims description 123
- 238000001514 detection method Methods 0.000 claims description 65
- 238000011176 pooling Methods 0.000 claims description 33
- 238000000034 method Methods 0.000 claims description 16
- 230000008569 process Effects 0.000 claims description 14
- 238000010606 normalization Methods 0.000 claims description 13
- 230000008859 change Effects 0.000 claims description 11
- 230000014509 gene expression Effects 0.000 claims description 8
- 210000002569 neuron Anatomy 0.000 claims description 8
- 239000013598 vector Substances 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 description 9
- 230000002159 abnormal effect Effects 0.000 description 7
- 230000004913 activation Effects 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image processing system based on personnel identification, which relates to the technical field of image processing, wherein a parameter updating module updates face detector parameters and feature extraction network parameters based on a gradient descent algorithm according to acquired image data and feedback information, a testing module tests newly acquired data according to the updated face detector parameters and feature extraction network parameters, an evaluation module evaluates the overall identification effect according to a test result, a judging module carries out corresponding processing according to the overall identification effect evaluation result, if the performance of the system is improved, the current parameters are continuously kept for use, and if the performance is not improved, the parameters are readjusted. The processing system dynamically adjusts algorithm parameters according to feedback information of the real-time scene to achieve the optimal recognition effect, so that the robustness and stability of the processing system are improved.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to an image processing system based on personnel identification.
Background
With the development of society and the advancement of technology, there is an increasing demand for security, and various places need to ensure that only authorized persons can enter to prevent unauthorized persons from entering an area where risks may occur, and a processing system is a system for implementing security access control using computer vision technology, which is generally used for monitoring places such as enterprise buildings, airports, subway stations, etc., to ensure that only authorized persons can enter a specific area.
The prior art has the following defects:
The existing processing system adopts fixed parameter setting, cannot adapt to the changes under different environments and scenes, and has large fluctuation of recognition performance caused by the influence of factors such as illumination, posture, shielding, multi-person crossing and the like in the actual application process, so that the robustness and stability of the processing system are poor;
Based on the above, the invention provides an image processing system based on personnel identification, which dynamically adjusts algorithm parameters according to feedback information of a real-time scene so as to realize the optimal identification effect, thereby improving the robustness and stability of the processing system.
Disclosure of Invention
The invention aims to provide an image processing system based on personnel identification, which aims to solve the defects in the background technology.
In order to achieve the above object, the present invention provides the following technical solutions: an image processing system based on personnel identification comprises an initialization module, a data acquisition module, a parameter updating module, a test module, an evaluation module and a judgment module;
An initialization module: initializing human face detector parameters and feature extraction network parameters;
and a data acquisition module: when the gate operates, continuously collecting image data when staff enter an office building, and simultaneously recording feedback information of each staff;
Parameter updating module: updating the parameters of the face detector and the network parameters extracted by the characteristics based on a gradient descent algorithm according to the acquired image data and feedback information;
and a testing module: testing the newly acquired data through the updated face detector parameters and the feature extraction network parameters;
And an evaluation module: evaluating the overall recognition effect of the processing system according to the test result;
And a judging module: and carrying out corresponding processing according to the overall recognition effect evaluation result, if the performance of the processing system is improved, keeping the current parameters continuously in use, and if the performance of the processing system is not improved, readjusting the parameters of the face detector and the parameters of the feature extraction network.
In a preferred embodiment, the initialization module initializes face detector parameters including a detection threshold, an image size, and feature extraction network parameters including a network layer number.
In a preferred embodiment, the parameter updating module obtains the false detection rate and the F1 score of the face detector in the feedback information, and then minimizes the false detection rate and maximizes the F1 score to generate the face detector loss function;
calculating a first partial derivative of the face detector loss function with respect to a face detector detection threshold using a back propagation algorithm, and updating the face detector detection threshold using a gradient descent algorithm based on the calculated first partial derivative;
Comparing the face detector loss function with a preset first iteration threshold, wherein the first iteration threshold is used for judging whether to stop updating the face detector detection threshold, applying the updated face detector detection threshold obtained each time to the face detector, judging that the face detector does not reach the expected level if the face detector loss function is larger than the first iteration threshold, continuing to update the face detector detection threshold, judging that the face detector reaches the expected level if the face detector loss function is smaller than or equal to the first iteration threshold, and stopping updating the face detector detection threshold.
In a preferred embodiment, the parameter updating module uses the cross entropy loss function as a feature extraction network loss function and updates the feature extraction network parameters based on a gradient descent algorithm;
calculating a second partial derivative of the feature extraction network loss function with respect to the feature extraction network parameter using a back propagation algorithm, the second partial derivative representing a rate of change of the feature extraction network loss function at the feature extraction network parameter;
optimizing and adjusting the characteristic extraction network parameters according to the comparison result of the second partial derivative and 0;
Comparing the feature extraction network loss function with a preset second iteration threshold, wherein the second iteration threshold is used for judging whether to stop updating the feature extraction network parameters, and applying the updated feature extraction network parameters to the feature extraction network;
If the feature extraction network loss function is larger than the second iteration threshold, judging that the feature extraction network does not reach the expected level, continuing to update the feature extraction network parameters, and if the feature extraction network loss function is smaller than or equal to the second iteration threshold, judging that the feature extraction network reaches the expected level, and stopping updating the feature extraction network parameters.
In a preferred embodiment, the feature extraction network parameters include a convolution layer, a pooling layer, a full connection layer;
if the second partial derivative of the convolution layer is smaller than 0, the convolution layer is indicated to be in negative correlation with the feature extraction network loss function, the convolution layer is deleted, if the second partial derivative of the convolution layer is equal to 0, the convolution layer is indicated to be not in correlation with the feature extraction network loss function, the number of convolution kernels in the convolution layer is reduced, and if the second partial derivative of the convolution layer is larger than 0, the convolution layer is indicated to be in positive correlation with the feature extraction network loss function, and the number of the convolution kernels in the convolution layer is increased;
If the second partial derivative of the pooling layer is smaller than 0, the pooling layer is inversely related to the feature extraction network loss function, the pooling layer is deleted, if the second partial derivative of the pooling layer is equal to 0, the pooling layer is irrelevant to the feature extraction network loss function, the size of the pooling layer is reduced, and if the second partial derivative of the pooling layer is larger than 0, the pooling layer is positively related to the feature extraction network loss function, and the size of the pooling layer is increased;
If the second partial derivative of the full-connection layer is smaller than 0, the full-connection layer is inversely related to the feature extraction network loss function, the full-connection layer is deleted, if the second partial derivative of the full-connection layer is equal to 0, the full-connection layer is irrelevant to the feature extraction network loss function, the number of neurons in the full-connection layer is reduced, and if the second partial derivative of the full-connection layer is larger than 0, the full-connection layer is positively related to the feature extraction network loss function, and the number of neurons in the full-connection layer is increased.
In a preferred embodiment, the parameter updating module obtains the false detection rate and the F1 score of the face detector in the feedback information, and then minimizes the false detection rate and maximizes the F1 score to generate the face detector loss function, where the expression is: In which, in the process, For the face detector loss function,In order to be able to detect the rate of false positives,The fraction of F1 is given as the fraction,To balance the false detection rate and the weight parameter of the F1 fraction, andThe value range is [0,1].
In a preferred embodiment, the parameter updating module uses a gradient descent algorithm to update the face detector detection threshold based on the calculated first partial derivativeThe updated formula is:
In which, in the process, To update the post-face detector detection threshold,To update the pre-face detector detection threshold,In order for the rate of learning to be high,As the first partial derivative, the first partial derivativeRepresenting face detector loss functionsAt the face detector detection thresholdRate of change at (c).
In a preferred embodiment, the expression of the feature extraction network loss function is:
In which, in the process, A network loss function is extracted for the feature,The input data is represented by a representation of the input data,Representing the corresponding real label(s),Representing feature extraction network parameters, including weights and biases,Representing classifier parameters, including weights and biases,Representing the output of the feature extraction network, i.e. the feature vector,The result of the prediction is indicated,For the number of categories to be considered,The representation is the true labelEach element is,Representing predicted firstProbability of individual categories.
In a preferred embodiment, the evaluation module obtains a face detector loss function in the test resultsFeature extraction network loss function；
Normalizing the face detector loss function and the feature extraction network loss function to map the value ranges of the face detector loss function and the feature extraction network loss function to be between [0,1], and adding the face detector loss function normalization value and the feature extraction network loss function normalization value to obtain a test score after obtaining the face detector loss function normalization value and the feature extraction network loss function normalization value;
And comparing the obtained test score with a score threshold, if the test score is smaller than or equal to the score threshold, evaluating that the overall recognition effect is good, and if the test score is larger than the score threshold, evaluating that the overall recognition effect is poor.
In a preferred embodiment, if the evaluation module evaluates that the overall recognition effect is good, the judgment module judges that the performance of the processing system is improved, and keeps the current face detector parameters and the feature extraction network parameters continuously used, and if the evaluation module evaluates that the overall recognition effect is poor, the judgment module judges that the performance of the processing system is not improved, and the face detector parameters and the feature extraction network parameters need to be readjusted.
In the technical scheme, the invention has the technical effects and advantages that:
1. According to the invention, the parameter updating module updates the face detector parameters and the feature extraction network parameters based on the gradient descent algorithm according to the acquired image data and the feedback information, the testing module tests the newly acquired data according to the updated face detector parameters and the feature extraction network parameters, the evaluation module evaluates the overall recognition effect according to the test result, the judgment module carries out corresponding processing according to the overall recognition effect evaluation result, if the performance of the system is improved, the current parameters are continuously kept for use, and if the performance is not improved, the parameters are readjusted. The processing system dynamically adjusts algorithm parameters according to feedback information of a real-time scene to achieve the optimal recognition effect, so that the robustness and stability of the processing system are improved;
2. According to the invention, when the parameters of the face detector and the parameters of the feature extraction network are synchronously updated, the judging module acquires the updating iteration times of the parameters of the face detector and the updating iteration times of the parameters of the feature extraction network in real time, obtains the abnormal coefficients after carrying out weighted calculation on the updating iteration times of the parameters of the face detector and the updating iteration times of the parameters of the feature extraction network, and when the parameters of the face detector and the parameters of the feature extraction network are synchronously updated, if the abnormal coefficients are larger than the preset abnormal threshold, the judging module judges that the equipment is likely to have faults, stops the iterative updating of the parameters of the face detector and the parameters of the feature extraction network, and sends warning signals to an administrator, so that the processing system is effectively prevented from entering a dead loop calculation state, and fault early warning can be carried out on the equipment.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments described in the present application, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.
FIG. 1 is a block diagram of a system according to the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1: referring to fig. 1, the image processing system based on personnel identification in this embodiment is used for identifying personnel entering an office building, and includes main steps of face detection, face alignment, feature extraction and classification identification, including an initialization module, a data acquisition module, a parameter update module, a test module, an evaluation module and a judgment module;
an initialization module: initializing face detector parameters and feature extraction network parameters, wherein the face detector parameters comprise detection threshold values, image sizes and the like, the feature extraction network parameters comprise network layers and the like, and the initialized face detector parameters and the feature extraction network parameters are sent to a parameter updating module;
and a data acquisition module: when the gate machine runs, continuously collecting image data when staff enter an office building, and simultaneously recording feedback information of each staff, wherein the feedback information comprises identity information and accuracy of a recognition result, and the image data and the feedback information are sent to a parameter updating module;
Parameter updating module: updating the parameters of the face detector and the parameters of the feature extraction network based on a gradient descent algorithm according to the acquired image data and feedback information, wherein the steps include adjusting the detection threshold value of the face detector and the parameters of the feature extraction network, and transmitting the updated parameters of the face detector and the updated parameters of the feature extraction network to a test module;
And a testing module: testing the newly acquired data through the updated face detector parameters and the updated feature extraction network parameters, and sending the test result to an evaluation module;
And an evaluation module: evaluating the overall recognition effect according to the test result, and transmitting the evaluation result to a judging module;
And a judging module: and carrying out corresponding processing according to the overall recognition effect evaluation result, if the performance of the system is improved, continuing to keep the current parameter in use, and if the performance is not improved, readjusting the parameter.
According to the application, the parameter updating module updates the face detector parameters and the feature extraction network parameters based on the gradient descent algorithm according to the acquired image data and the feedback information, the testing module tests the newly acquired data according to the updated face detector parameters and the feature extraction network parameters, the evaluation module evaluates the overall recognition effect according to the test result, the judgment module carries out corresponding processing according to the overall recognition effect evaluation result, if the performance of the system is improved, the current parameters are continuously kept for use, and if the performance is not improved, the parameters are readjusted. The processing system dynamically adjusts algorithm parameters according to feedback information of the real-time scene to achieve the optimal recognition effect, so that the robustness and stability of the processing system are improved.
The specific working flow is as follows:
The processing system initializes face detector parameters and feature extraction network parameters, the face detector parameters comprise detection threshold values, image sizes and the like, the feature extraction network parameters comprise network layers and the like, when a gate machine operates, image data of staff entering an office building is continuously collected, meanwhile, the identity information of each staff and the accuracy of recognition results are recorded, the face detector parameters and the feature extraction network parameters are updated according to the collected image data and feedback information based on a gradient descent algorithm, the face detector detection threshold values and the feature extraction network parameters comprise adjustment of the face detector detection threshold values and the feature extraction network parameters, the newly collected data are tested through the updated face detector parameters and the feature extraction network parameters, the overall recognition effect of the processing system is evaluated according to the test results, corresponding processing is carried out according to the overall recognition effect evaluation results, if the performance of the system is improved, the current parameter use is continuously kept, and if the performance is not improved, the parameters are readjusted.
An initialization module: initializing face detector parameters and feature extraction network parameters, wherein the face detector parameters comprise detection threshold values, image sizes and the like, and the feature extraction network parameters comprise network layers and the like;
initializing face detector parameters: an initial detection threshold is determined, which is a confidence threshold, typically a probability value or score value, for determining the detected face. The initial image size is determined, which refers to the size of the image input to the face detector, typically in pixels.
Initializing feature extraction network parameters: the architecture of the network is determined, including the number of layers of the network, the number of nodes per layer, the activation function, etc. The weights and biases of the network are initialized, and random initialization or pre-trained weights may be used.
For example:
Initializing face detector parameters: the detection threshold is set to an initial threshold of 0.5, which means that only if the probability of detecting a face is greater than 0.5, it is considered as a valid detection result. The initial image size is determined to be 224x224 pixels.
Initializing feature extraction network parameters: a pre-trained convolutional neural network is selected as a feature extraction network, such as VGG, resNet, etc., and pre-trained weights are used to initialize the network, or random initialization is performed.
And a data acquisition module: when the gate operates, continuously collecting image data when staff enter an office building, and simultaneously recording feedback information of each staff, wherein the feedback information comprises identity information and accuracy of an identification result;
When the gate operates, image data of staff entering an office building is continuously acquired from a camera or other image acquisition equipment. Ensuring that the acquired image is of sufficient quality and resolution for subsequent personnel identification tasks.
And recording feedback information of staff, including identity information and accuracy of recognition results, of each acquired image. The identity information may be the employee's job number, name, etc. for identifying the employee's identity. The Accuracy of the recognition result can be used to evaluate the performance of the system, typically expressed in terms of Accuracy (Accuracy). The acquired image data and feedback information are stored in a database or file system.
Parameter updating module: updating the parameters of the face detector and the parameters of the feature extraction network based on a gradient descent algorithm according to the acquired image data and feedback information, wherein the updating comprises the steps of adjusting the detection threshold value of the face detector and the parameters of the feature extraction network;
the parameter updating module acquires the false detection rate and F1 fraction of the face detector in the feedback information, and then minimizes the false detection rate and maximizes the F1 fraction to generate a face detector loss function, wherein the expression is as follows: In which, in the process, For the face detector loss function,In order to be able to detect the rate of false positives,The fraction of F1 is given as the fraction,To balance the false detection rate and the weight parameter of the F1 fraction, andThe value range is [0,1], whenNear 0, the loss function is more concerned with maximizing the F1 score whenApproaching 1, the loss function is more concerned with minimizing the false detection rate.
Computing a first partial derivative of a face detector loss function with respect to a face detector detection threshold using a back propagation algorithmFirst partial derivativeRepresenting face detector loss functionsAt the face detector detection thresholdRate of change at;
if the first partial derivative is greater than 0, this means that the face detector detection threshold is increased Can lead to a face detector loss functionIncreasing and decreasing the face detector detection thresholdCan lead to a face detector loss functionA reduction;
If the first partial derivative is less than 0, this means that the face detector detection threshold is increased Can lead to a face detector loss functionReducing the face detector detection thresholdCan lead to a face detector loss functionIncreasing;
if the first partial derivative is equal to 0, this means that the threshold is detected at the face detector Loss function of face detectorIs zero, i.e. face detector loss functionAt which point a local extremum is reached.
Updating the face detector detection threshold using a gradient descent algorithm based on the calculated first partial derivativeThe updated formula is:
In which, in the process, To update the post-face detector detection threshold,To update the pre-face detector detection threshold,In order for the rate of learning to be high,Is the first partial derivative;
Comparing the face detector loss function with a preset first iteration threshold, wherein the first iteration threshold is used for judging whether to stop updating the face detector detection threshold, applying the updated face detector detection threshold obtained each time to the face detector, judging that the face detector does not reach the expected level if the face detector loss function is larger than the first iteration threshold, continuing to update the face detector detection threshold, judging that the face detector reaches the expected level if the face detector loss function is smaller than or equal to the first iteration threshold, and stopping updating the face detector detection threshold.
For example, assume that the initial face detector detection threshold is 0.6 and the learning rate is 0.01. By the gradient descent algorithm, the first partial derivative calculated by the user is-0.002, which means that the user needs to adjust the detection threshold of the face detector by a small step size upwards to reduce the loss function of the face detector. Therefore, we update the face detector detection threshold to 0.602 and continue iterating this process until the face detector loss function is less than or equal to the first iteration threshold and the face detector performance reaches the desired level.
In the application, the feature extraction network performs image classification tasks through a Convolutional Neural Network (CNN), and the network structure is as follows:
An input layer;
Convolution layer (Convolutional Layer): 32 3x3 convolution kernels, reLU activation function;
Pooling layer (Pooling Layer): 2x2 max pooling layer;
Full tie layer (Fully Connected Layer): 128 neurons, reLU activation function;
Output layer: 10 neurons, softmax activation function (applicable to class 10 classification tasks);
The parameter updating module uses the cross entropy loss function as a feature extraction network loss function and updates the feature extraction network parameters based on a gradient descent algorithm;
the expression of the feature extraction network loss function is as follows:
In which, in the process, A network loss function is extracted for the feature,The input data is represented by a representation of the input data,Representing the corresponding real label(s),Representing feature extraction network parameters, including weights and biases,Representing classifier parameters, including weights and biases,Representing the output of the feature extraction network, i.e. the feature vector,The result of the prediction is indicated,For the number of categories to be considered,The representation is the true labelEach element is,Representing predicted firstProbability of individual category;
The feature extraction network loss function is used for measuring the difference between the predicted value and the real label, namely the fitting degree of the training sample. When the value of the feature extraction network loss function is smaller, the difference between the predicted result and the real label is smaller, and the fitting degree on the training data is better. Therefore, the objective of the optimization process is to minimize the feature extraction network loss function so that the face recognition model performs better on the training data.
Calculating a second partial derivative of the feature extraction network loss function with respect to the feature extraction network parameters using a back propagation algorithmSecond partial derivativeRepresenting feature extraction network loss functionsExtracting network parameters at featuresRate of change at;
The characteristic extraction network parameters comprise a convolution layer, a pooling layer and a full connection layer;
if the second partial derivative of the convolution layer is smaller than 0, the convolution layer is indicated to be in negative correlation with the feature extraction network loss function, the convolution layer is deleted, if the second partial derivative of the convolution layer is equal to 0, the convolution layer is indicated to be not in correlation with the feature extraction network loss function, the number of convolution kernels in the convolution layer is reduced, and if the second partial derivative of the convolution layer is larger than 0, the convolution layer is indicated to be in positive correlation with the feature extraction network loss function, and the number of the convolution kernels in the convolution layer is increased;
If the second partial derivative of the pooling layer is smaller than 0, the pooling layer is inversely related to the feature extraction network loss function, the pooling layer is deleted, if the second partial derivative of the pooling layer is equal to 0, the pooling layer is irrelevant to the feature extraction network loss function, the size of the pooling layer is reduced, and if the second partial derivative of the pooling layer is larger than 0, the pooling layer is positively related to the feature extraction network loss function, and the size of the pooling layer is increased;
If the second partial derivative of the full-connection layer is smaller than 0, the full-connection layer is inversely related to the feature extraction network loss function, the full-connection layer is deleted, if the second partial derivative of the full-connection layer is equal to 0, the full-connection layer is irrelevant to the feature extraction network loss function, the number of neurons in the full-connection layer is reduced, and if the second partial derivative of the full-connection layer is larger than 0, the full-connection layer is positively related to the feature extraction network loss function, and the number of neurons in the full-connection layer is increased.
According to the chain law, the second partial derivative is decomposed into:；
Wherein;
Calculation of ：Representing feature extraction network loss functionsRegarding the prediction resultThe change rate of the network loss function can be directly calculated by the definition of the characteristic extraction network loss function:。
Calculation of ：Representing predicted resultsRegarding the rate of change of the feature extraction network output, it is calculated by back-propagation of the classifier:
；
In the method, in the process of the invention, Refers to the step of calculating the rate of change of the classifier (or fully connected layer) during the back propagation.
Calculation of：Representing the rate of change of the feature extraction network output with respect to the feature extraction network parameters, calculated by back-propagation of the feature extraction network:
；
In the method, in the process of the invention, Refers to the step of calculating the rate of change of the feature extraction network (or feature extractor) during the back propagation process.
Comparing the feature extraction network loss function with a preset second iteration threshold, wherein the second iteration threshold is used for judging whether to stop updating the feature extraction network parameters, the updated feature extraction network parameters are applied to the feature extraction network, if the feature extraction network loss function is larger than the second iteration threshold, the feature extraction network is judged not to reach the expected level, the feature extraction network parameters are continuously updated, and if the feature extraction network loss function is smaller than or equal to the second iteration threshold, the feature extraction network is judged to reach the expected level, and the feature extraction network parameters are stopped being updated.
And a testing module: testing the newly acquired data through the updated face detector parameters and the feature extraction network parameters;
Acquiring newly acquired image data and corresponding labels, and preprocessing the image data, such as image normalization, size adjustment and the like, so as to keep the consistency of the image data;
And reasoning the test data by using the updated face detector parameters and the feature extraction network parameters. For a face detection task, a face position and a bounding box in an image need to be detected; for a feature extraction network, feature vectors in an image need to be extracted.
And an evaluation module: evaluating the overall recognition effect according to the test result;
Obtaining a face detector loss function in a test result Feature extraction network loss function；
Normalizing the face detector loss function and the feature extraction network loss function to map the value ranges of the face detector loss function and the feature extraction network loss function to be between [0,1], and adding the face detector loss function normalization value and the feature extraction network loss function normalization value to obtain a test score after obtaining the face detector loss function normalization value and the feature extraction network loss function normalization value;
The expressions of the face detector loss function and the feature extraction network loss function show that the smaller the values are, the better the face detector loss function and the feature extraction network loss function are;
Therefore, the smaller the obtained test score is, the better the overall recognition effect is, the obtained test score is compared with a preset score threshold, if the test score is smaller than or equal to the score threshold, the overall recognition effect is estimated to be good, and if the test score is larger than the score threshold, the overall recognition effect is estimated to be poor.
And a judging module: performing corresponding processing according to the overall recognition effect evaluation result, if the performance of the system is improved, continuously keeping the current parameters for use, and if the performance is not improved, readjusting the parameters;
If the evaluation module evaluates that the overall recognition effect is good, judging that the performance of the processing system is improved, continuously keeping the current face detector parameters and the feature extraction network parameters to be used, if the evaluation module evaluates that the overall recognition effect is poor, judging that the performance of the processing system is not improved, and readjusting the face detector parameters and the feature extraction network parameters.
Example 2: in practical application, if the face recognition error is large and the efficiency is low due to the equipment failure, even if the face detector parameters and the feature extraction network parameters are updated and optimized continuously, the convergence condition is still not met (namely, the face detector loss function is still greater than the first iteration threshold after the face detector parameters are updated for many times, and the feature extraction network parameters are still greater than the second iteration threshold after the feature extraction network parameters are updated for many times), and the updating and optimization of the face detector parameters and the feature extraction network parameters are performed synchronously, if the equipment failure is not detected by the gate, the processing system is caused to perform iterative computation continuously, so that the redundant data quantity is increased, and the calculation load of the processing system is increased;
Therefore, in order to avoid the above situation, when the parameters of the face detector and the parameters of the feature extraction network are updated synchronously, the judging module acquires the number of iterations of updating the parameters of the face detector and the number of iterations of updating the parameters of the feature extraction network in real time, and obtains the abnormal coefficients after weighting calculation of the number of iterations of updating the parameters of the face detector and the number of iterations of updating the parameters of the feature extraction network, wherein the expression is as follows: In which, in the process, As the coefficient of anomaly it is,The iteration number is updated for the face detector parameters,Extracting network parameters for the features and updating iteration times;
when the parameters of the face detector and the parameters of the feature extraction network are synchronously updated, if the anomaly coefficient is larger than a preset anomaly threshold value, the judging module judges that the equipment possibly has faults, stops iterative updating of the parameters of the face detector and the parameters of the feature extraction network, and sends a warning signal to an administrator;
According to the application, when the parameters of the face detector and the parameters of the feature extraction network are synchronously updated, the judging module acquires the updating iteration times of the parameters of the face detector and the updating iteration times of the parameters of the feature extraction network in real time, obtains the abnormal coefficients after carrying out weighted calculation on the updating iteration times of the parameters of the face detector and the updating iteration times of the parameters of the feature extraction network, and when the parameters of the face detector and the parameters of the feature extraction network are synchronously updated, if the abnormal coefficients are larger than the preset abnormal threshold, the judging module judges that the equipment is likely to have faults, stops the iterative updating of the parameters of the face detector and the parameters of the feature extraction network, and sends warning signals to an administrator, so that the processing system is effectively prevented from entering a dead loop calculation state, and fault early warning can be carried out on the equipment.
The above formulas are all formulas with dimensions removed and numerical values calculated, the formulas are formulas with a large amount of data collected for software simulation to obtain the latest real situation, and preset parameters in the formulas are set by those skilled in the art according to the actual situation.
In the description of the present specification, the descriptions of the terms "one embodiment," "example," "specific example," and the like, mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The preferred embodiments of the invention disclosed above are intended only to assist in the explanation of the invention. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise form disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best understand and utilize the invention. The invention is limited only by the claims and the full scope and equivalents thereof.
Claims (10)
1. An image processing system based on personnel identification, characterized in that: the system comprises an initialization module, a data acquisition module, a parameter updating module, a test module, an evaluation module and a judgment module;
An initialization module: initializing human face detector parameters and feature extraction network parameters;
and a data acquisition module: when the gate operates, continuously collecting image data when staff enter an office building, and simultaneously recording feedback information of each staff;
Parameter updating module: updating the parameters of the face detector and the network parameters extracted by the characteristics based on a gradient descent algorithm according to the acquired image data and feedback information;
and a testing module: testing the newly acquired data through the updated face detector parameters and the feature extraction network parameters;
And an evaluation module: evaluating the overall recognition effect of the processing system according to the test result;
And a judging module: and carrying out corresponding processing according to the overall recognition effect evaluation result, if the performance of the processing system is improved, keeping the current parameters continuously in use, and if the performance of the processing system is not improved, readjusting the parameters of the face detector and the parameters of the feature extraction network.
2. An image processing system based on person identification as claimed in claim 1, wherein: the initialization module initializes face detector parameters including detection threshold and image size and feature extraction network parameters including network layer number.
3. An image processing system based on person identification as claimed in claim 2, wherein: the parameter updating module acquires the false detection rate and F1 fraction of the face detector in the feedback information, and then minimizes the false detection rate and maximizes the F1 fraction to generate a face detector loss function;
calculating a first partial derivative of the face detector loss function with respect to a face detector detection threshold using a back propagation algorithm, and updating the face detector detection threshold using a gradient descent algorithm based on the calculated first partial derivative;
Comparing the face detector loss function with a preset first iteration threshold, wherein the first iteration threshold is used for judging whether to stop updating the face detector detection threshold, applying the updated face detector detection threshold obtained each time to the face detector, judging that the face detector does not reach the expected level if the face detector loss function is larger than the first iteration threshold, continuing to update the face detector detection threshold, judging that the face detector reaches the expected level if the face detector loss function is smaller than or equal to the first iteration threshold, and stopping updating the face detector detection threshold.
4. An image processing system based on person identification as claimed in claim 2, wherein: the parameter updating module uses the cross entropy loss function as a feature extraction network loss function and updates feature extraction network parameters based on a gradient descent algorithm;
calculating a second partial derivative of the feature extraction network loss function with respect to the feature extraction network parameter using a back propagation algorithm, the second partial derivative representing a rate of change of the feature extraction network loss function at the feature extraction network parameter;
optimizing and adjusting the characteristic extraction network parameters according to the comparison result of the second partial derivative and 0;
Comparing the feature extraction network loss function with a preset second iteration threshold, wherein the second iteration threshold is used for judging whether to stop updating the feature extraction network parameters, and applying the updated feature extraction network parameters to the feature extraction network;
If the feature extraction network loss function is larger than the second iteration threshold, judging that the feature extraction network does not reach the expected level, continuing to update the feature extraction network parameters, and if the feature extraction network loss function is smaller than or equal to the second iteration threshold, judging that the feature extraction network reaches the expected level, and stopping updating the feature extraction network parameters.
5. An image processing system based on person identification as claimed in claim 4, wherein: the characteristic extraction network parameters comprise a convolution layer, a pooling layer and a full connection layer;
if the second partial derivative of the convolution layer is smaller than 0, the convolution layer is indicated to be in negative correlation with the feature extraction network loss function, the convolution layer is deleted, if the second partial derivative of the convolution layer is equal to 0, the convolution layer is indicated to be not in correlation with the feature extraction network loss function, the number of convolution kernels in the convolution layer is reduced, and if the second partial derivative of the convolution layer is larger than 0, the convolution layer is indicated to be in positive correlation with the feature extraction network loss function, and the number of the convolution kernels in the convolution layer is increased;
If the second partial derivative of the pooling layer is smaller than 0, the pooling layer is inversely related to the feature extraction network loss function, the pooling layer is deleted, if the second partial derivative of the pooling layer is equal to 0, the pooling layer is irrelevant to the feature extraction network loss function, the size of the pooling layer is reduced, and if the second partial derivative of the pooling layer is larger than 0, the pooling layer is positively related to the feature extraction network loss function, and the size of the pooling layer is increased;
If the second partial derivative of the full-connection layer is smaller than 0, the full-connection layer is inversely related to the feature extraction network loss function, the full-connection layer is deleted, if the second partial derivative of the full-connection layer is equal to 0, the full-connection layer is irrelevant to the feature extraction network loss function, the number of neurons in the full-connection layer is reduced, and if the second partial derivative of the full-connection layer is larger than 0, the full-connection layer is positively related to the feature extraction network loss function, and the number of neurons in the full-connection layer is increased.
6. An image processing system based on person identification as claimed in claim 3, characterized in that: the parameter updating module acquires the false detection rate and F1 fraction of the face detector in the feedback information, and then minimizes the false detection rate and maximizes the F1 fraction to generate a loss function of the face detector, wherein the expression is as follows: In which, in the process, For the face detector loss function,/>Is the false detection rate,/>Is F1 score,/>To balance the false detection rate and the weight parameter of the F1 fraction, and/>The value range is [0,1].
7. An image processing system based on person identification as claimed in claim 3, characterized in that: the parameter updating module uses a gradient descent algorithm to update the detection threshold of the face detector according to the calculated first partial derivativeThe updated formula is:
In the above, the ratio of/> For updated face detector detection threshold,/>To update the pre-face detector detection threshold,/>In order for the rate of learning to be high,As the first partial derivative, the first partial derivative/>Representing the face detector loss function/>Detection threshold at face detector/>Rate of change at (c).
8. An image processing system based on person identification as claimed in claim 4, wherein: the expression of the feature extraction network loss function is as follows:
In the above, the ratio of/> Extracting a network loss function for a feature,/>Representing input data,/>Representing the corresponding real label,/>Representing feature extraction network parameters including weights and biases,/>Representing classifier parameters including weights and biases,/>Representing the output of the feature extraction network, i.e. feature vectors,/>Representing the predicted result,/>Is the category number,/>The representation is the true label-Individual element,/>Represents predicted first/>Probability of individual categories.
9. An image processing system based on person identification as claimed in claim 4, wherein: the evaluation module obtains a loss function of the face detector in the test resultFeature extraction network loss function/>；
Normalizing the face detector loss function and the feature extraction network loss function to map the value ranges of the face detector loss function and the feature extraction network loss function to be between [0,1], and adding the face detector loss function normalization value and the feature extraction network loss function normalization value to obtain a test score after obtaining the face detector loss function normalization value and the feature extraction network loss function normalization value;
And comparing the obtained test score with a score threshold, if the test score is smaller than or equal to the score threshold, evaluating that the overall recognition effect is good, and if the test score is larger than the score threshold, evaluating that the overall recognition effect is poor.
10. An image processing system based on person identification as claimed in claim 4, wherein: if the evaluation module evaluates that the overall recognition effect is good, the judgment module judges that the performance of the processing system is improved, the current face detector parameters and the feature extraction network parameters are kept continuously in use, and if the evaluation module evaluates that the overall recognition effect is poor, the judgment module judges that the performance of the processing system is not improved, and the face detector parameters and the feature extraction network parameters need to be readjusted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410533876.6A CN118116061B (en) | 2024-04-30 | 2024-04-30 | Image processing system based on personnel identification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410533876.6A CN118116061B (en) | 2024-04-30 | 2024-04-30 | Image processing system based on personnel identification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118116061A true CN118116061A (en) | 2024-05-31 |
| CN118116061B CN118116061B (en) | 2024-10-18 |
Family
ID=91212720
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410533876.6A Active CN118116061B (en) | 2024-04-30 | 2024-04-30 | Image processing system based on personnel identification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118116061B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446689A (en) * | 2018-05-30 | 2018-08-24 | 南京开为网络科技有限公司 | A kind of face identification method |
| CN110246244A (en) * | 2019-05-16 | 2019-09-17 | 珠海华园信息技术有限公司 | Intelligent foreground management system based on recognition of face |
| CN110245550A (en) * | 2019-04-22 | 2019-09-17 | 北京云识图信息技术有限公司 | A kind of face noise data collection CNN training method based on overall cosine distribution |
| CN110458133A (en) * | 2019-08-19 | 2019-11-15 | 电子科技大学 | Lightweight method for detecting human face based on production confrontation network |
| WO2020102988A1 (en) * | 2018-11-20 | 2020-05-28 | 西安电子科技大学 | Feature fusion and dense connection based infrared plane target detection method |
| CN111222434A (en) * | 2019-12-30 | 2020-06-02 | 深圳市爱协生科技有限公司 | Method for obtaining evidence of synthesized face image based on local binary pattern and deep learning |
| CN112364803A (en) * | 2020-11-20 | 2021-02-12 | 深圳龙岗智能视听研究院 | Living body recognition auxiliary network and training method, terminal, equipment and storage medium |
| CN114692741A (en) * | 2022-03-21 | 2022-07-01 | 华南理工大学 | Generalized face counterfeiting detection method based on domain invariant features |
| CN116385969A (en) * | 2023-04-07 | 2023-07-04 | 暨南大学 | Personnel gathering detection system based on multi-camera cooperation and human feedback |
-
2024
- 2024-04-30 CN CN202410533876.6A patent/CN118116061B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446689A (en) * | 2018-05-30 | 2018-08-24 | 南京开为网络科技有限公司 | A kind of face identification method |
| WO2020102988A1 (en) * | 2018-11-20 | 2020-05-28 | 西安电子科技大学 | Feature fusion and dense connection based infrared plane target detection method |
| CN110245550A (en) * | 2019-04-22 | 2019-09-17 | 北京云识图信息技术有限公司 | A kind of face noise data collection CNN training method based on overall cosine distribution |
| CN110246244A (en) * | 2019-05-16 | 2019-09-17 | 珠海华园信息技术有限公司 | Intelligent foreground management system based on recognition of face |
| CN110458133A (en) * | 2019-08-19 | 2019-11-15 | 电子科技大学 | Lightweight method for detecting human face based on production confrontation network |
| CN111222434A (en) * | 2019-12-30 | 2020-06-02 | 深圳市爱协生科技有限公司 | Method for obtaining evidence of synthesized face image based on local binary pattern and deep learning |
| CN112364803A (en) * | 2020-11-20 | 2021-02-12 | 深圳龙岗智能视听研究院 | Living body recognition auxiliary network and training method, terminal, equipment and storage medium |
| CN114692741A (en) * | 2022-03-21 | 2022-07-01 | 华南理工大学 | Generalized face counterfeiting detection method based on domain invariant features |
| CN116385969A (en) * | 2023-04-07 | 2023-07-04 | 暨南大学 | Personnel gathering detection system based on multi-camera cooperation and human feedback |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118116061B (en) | 2024-10-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111209434B (en) | Substation equipment inspection system and method based on multi-source heterogeneous data fusion | |
| CN110717481B (en) | Method for realizing face detection by using cascaded convolutional neural network | |
| CN110213244A (en) | A kind of network inbreak detection method based on space-time characteristic fusion | |
| CN111611874B (en) | Face mask wearing detection method based on ResNet and Canny | |
| CN113449660B (en) | Abnormal event detection method of space-time variation self-coding network based on self-attention enhancement | |
| CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
| CN109255441A (en) | Spacecraft fault diagnosis method based on artificial intelligence | |
| CN109858389A (en) | Vertical ladder demographic method and system based on deep learning | |
| CN111209832B (en) | Auxiliary obstacle avoidance training method, equipment and medium for substation inspection robot | |
| CN113869162A (en) | Violation identification method and system based on artificial intelligence | |
| CN113392931A (en) | Hyperspectral open set classification method based on self-supervision learning and multitask learning | |
| CN117809164A (en) | Substation equipment fault detection method and system based on multi-mode fusion | |
| CN112132005A (en) | Face detection method based on cluster analysis and model compression | |
| CN112149962A (en) | Risk quantitative evaluation method and system for cause behavior of construction accident | |
| CN116824335A (en) | YOLOv5 improved algorithm-based fire disaster early warning method and system | |
| CN113065431B (en) | Human body violation prediction method based on hidden Markov model and recurrent neural network | |
| CN117636477A (en) | Multi-target tracking matching method based on radial basis function fuzzy neural network | |
| CN112488213A (en) | Fire picture classification method based on multi-scale feature learning network | |
| CN114913485A (en) | Multi-level feature fusion weak supervision detection method | |
| CN113378638B (en) | Method for identifying abnormal behavior of turbine operator based on human body joint point detection and D-GRU network | |
| CN118116061B (en) | Image processing system based on personnel identification | |
| KR20230060214A (en) | Apparatus and Method for Tracking Person Image Based on Artificial Intelligence | |
| CN113935413A (en) | Distribution network wave recording file waveform identification method based on convolutional neural network | |
| CN118038021A (en) | Transformer substation operation site foreign matter intrusion detection method based on improvement yolov4 | |
| CN117877189A (en) | Overhead line channel mountain fire hidden danger identification method based on semi-supervised learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |