CN112329610B - High-voltage line detection method based on edge attention mechanism fusion network - Google Patents
High-voltage line detection method based on edge attention mechanism fusion network Download PDFInfo
- Publication number
- CN112329610B CN112329610B CN202011210495.2A CN202011210495A CN112329610B CN 112329610 B CN112329610 B CN 112329610B CN 202011210495 A CN202011210495 A CN 202011210495A CN 112329610 B CN112329610 B CN 112329610B
- Authority
- CN
- China
- Prior art keywords
- image
- voltage line
- network
- edge
- output
- 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.)
- Active
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 55
- 238000001514 detection method Methods 0.000 title claims abstract description 52
- 230000007246 mechanism Effects 0.000 title claims abstract description 24
- 238000012545 processing Methods 0.000 claims abstract description 42
- 238000000034 method Methods 0.000 claims abstract description 15
- 230000008569 process Effects 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 16
- 238000000605 extraction Methods 0.000 claims description 11
- 238000012952 Resampling Methods 0.000 claims description 8
- 238000005070 sampling Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000007499 fusion processing Methods 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 abstract description 14
- 238000005452 bending Methods 0.000 abstract description 6
- 239000000284 extract Substances 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 18
- 238000013135 deep learning Methods 0.000 description 12
- 230000006870 function Effects 0.000 description 8
- 238000013136 deep learning model Methods 0.000 description 7
- 238000010801 machine learning Methods 0.000 description 4
- 238000003708 edge detection Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 229910052704 radon Inorganic materials 0.000 description 1
- SYUHGPGVQRZVTB-UHFFFAOYSA-N radon atom Chemical compound [Rn] SYUHGPGVQRZVTB-UHFFFAOYSA-N 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a high-voltage line detection method based on an edge attention mechanism fusion network, which comprises the following steps: 1) The network reads the locally stored parameter weight; 2) The network automatically processes the input high-voltage line image to be detected, divides the high-voltage line image into two processing paths, extracts characteristic information of the image by using a convolution module, and then performs fusion operation by using a characteristic fusion module; 3) After the feature fusion module processes, the feature images are subjected to bilinear interpolation operation to restore to the original image size, then two-classification operation is performed, finally a binary predictive label image is output, and whether each pixel of the image belongs to a high-voltage line is judged. The invention effectively solves the problem of inaccurate high-voltage line detection of the semantic segmentation network, and can accurately detect the high-voltage line position information in the image; moreover, the adopted semantic segmentation network can effectively solve the detection problem of the bending high-voltage line.
Description
Technical Field
The invention relates to a high-voltage line detection method, in particular to a high-voltage line detection method based on an edge attention mechanism fusion network.
Background
At present, with the gradual opening of a low-altitude airspace, helicopters and unmanned aerial vehicles for low-altitude operation are widely focused and applied in various industry fields; particularly, as demands for executing tasks under low-altitude complex environments of helicopters and unmanned aerial vehicles are continuously increased, safety demands for guaranteeing the low-altitude complex environments of helicopters and unmanned aerial vehicles are also becoming more and more urgent. Among various obstacles, the high-voltage line is higher in position and fine in size, and the existing detection facilities are difficult to find, so that the high-voltage line is one of the most dangerous obstacles in low-altitude flight. The number of helicopters destroyed by a collision accident with a high voltage line is reported to be much greater than the number destroyed in war. Therefore, the rapid and accurate detection of the high-voltage line obstacle has important significance for the helicopter and the unmanned aerial vehicle to lift the early warning distance, reserving enough obstacle avoidance time and avoiding the discovery of collision accidents.
A method commonly used to help a helicopter or unmanned aerial vehicle avoid colliding with a high voltage line is to install an aviation warning ball for the high voltage line. Although the aviation warning ball can help the aircraft to detect the high-voltage line, the installation steps are tedious and time-consuming, the aviation warning ball cannot be used for installing aviation warners for all the high-voltage lines at all, and the aviation warning ball can only play a role in warning and cannot accurately detect the positions of the high-voltage lines. Meanwhile, the conventional high-voltage line detection method usually uses the high-voltage line as a straight line for detection by default, but due to the problems of gravity and the like, the high-voltage line is bent, so that the performance of the conventional high-voltage line detection method is poor. Therefore, it is necessary to develop a rapid and efficient accurate detection method for high voltage lines.
Disclosure of Invention
In order to solve the defects of the technology, the invention provides a high-voltage line detection method based on an edge attention mechanism fusion network.
In order to solve the technical problems, the invention adopts the following technical scheme: a high-voltage line detection method based on an edge attention mechanism fusion network comprises the following steps:
1) Network reads the locally stored parameter weights
After the network is subjected to supervision training, parameters are fixed, parameter weight model data are stored locally, and when high-voltage line detection is needed, the network reads the parameter weights from the local parameter weight model data again and loads the parameter weights into the network;
2) Inputting a high-voltage line image to be detected
Inputting a high-voltage line image to be detected to a network, and automatically processing the image by the network, wherein the processing operation of the network is specifically as follows: dividing the image into two processing paths, extracting characteristic information of the image by using a convolution module, and then carrying out fusion operation on the extracted deep semantic information and shallow space position information by using a characteristic fusion module;
3) Outputting a binary predictive label graph
Because the convolution module reduces the size of the image when extracting the characteristic information of the image, after the characteristic fusion module processes, the characteristic image is restored to the original image size by bilinear interpolation operation, then two-classification operation is carried out, whether each pixel of the image belongs to a high-voltage line is judged, and finally a binary predictive label image is output.
Further, for the input image to be detected, the specific processing procedures of the two processing paths are as follows:
The upper processing path is to perform three continuous downsampling convolution operations on an input image to be detected, each convolution module can enable the size of the input image to be half of the original size, after the three downsampling convolution operations, the size of an output feature image is 1/8 of the size of the original input image, and the path is used for retaining shallow space position information of the image;
The lower processing path is to perform five continuous downsampling convolution operations on the input image to be detected, and similarly, the size of the output feature image is 1/32 of that of the original input image, and the path is used for increasing the receptive field of the network, so that the network obtains higher-level semantic information, and the higher-level semantic information represents deep semantic information.
Further, the output feature graphs of the upper and lower processing paths are input to a feature fusion module, and the feature fusion module uses a pixel-level fusion strategy to perform fusion processing, and the specific operation is as follows: and (3) downsampling the 1/8 feature map output by the upper convolution module into a 1/32 feature map, performing Sigmoid activation on the 1/32 feature map output by the lower convolution module, and finally performing pixel-level multiplication operation with the 1/32 feature map above.
Further, the output of the feature fusion module is also divided into two paths, specifically:
The lower output path is to restore the 1/32 feature map to the original input image size by continuous up-sampling operation, then to perform two-classification operation to obtain the predicted main body label output by the model, and calculate the loss value with the real high-voltage line main body label to obtain the main body loss;
the upper output path is processed by an edge information extraction and fusion module, the upper output path passing through the three downsampling convolution modules carries out resampling operation on the output feature images, resampling feature images are obtained, then subtraction operation is carried out on the input feature images and the resampling feature images, edge information feature images are extracted, and edge information fusion is carried out on the input feature images and shallow feature images output by the convolution modules; then up-sampling operation is carried out to restore the original input image size, and finally classification operation is carried out to obtain a predicted edge label output by the model, and a loss value is calculated with the real high-voltage line edge label to obtain edge loss; wherein the shallow feature map is a feature map of the upper processing path output.
Further, after the main body loss and the edge loss are obtained, the main body loss and the edge loss are added to obtain total loss, and then the gradient descent algorithm is used for continuously reducing the total loss, so that the training process of the network is carried out, and the high-voltage line position information in the image is accurately detected.
The invention discloses a high-voltage line detection method based on an edge attention mechanism fusion network, which aims at the problem of inaccurate detection of a prediction label boundary of a semantic segmentation model of a deep learning algorithm, an edge attention mechanism module is added in the structure of the network, edge loss is also considered in a training optimization grid, the problem of inaccurate detection of the high-voltage line of the semantic segmentation network is effectively solved, and the position information of the high-voltage line in an image can be accurately detected; moreover, the adopted semantic segmentation network can effectively solve the detection problem of the bending high-voltage line.
Drawings
FIG. 1 is a diagram of a converged network architecture based on an edge attention mechanism in accordance with the present invention.
FIG. 2 is a flow chart of edge extraction information according to the present invention.
FIG. 3 is a graph showing comparison of results according to an embodiment of the present invention.
(A) Is an input image; (b) is a true label drawing; (c) predicting a label graph for a general semantic segmentation network; (d) To fuse network predictive label graphs based on edge attention mechanisms.
Detailed Description
The invention will be described in further detail with reference to the drawings and the detailed description.
The invention discloses a high-voltage line detection method based on an edge attention mechanism fusion network, aiming at the problem of inaccurate boundary detection of a prediction label of a semantic segmentation model of a deep learning algorithm, an edge attention mechanism module is added in the structure of the network, the calculation loss of the prediction edge label and a real edge label output by the edge attention mechanism module is calculated, the loss of the prediction label and the real label output by the model finally is added, a gradient descent method is used for training and learning the model together, the loss values of the prediction edge label and the real edge label are minimized as much as possible, and the model can be combined with the predicted edge label information to accurately detect the high-voltage line position information in an image.
The definition of semantic segmentation in the deep learning algorithm is to classify every pixel of the image, and pixels belonging to the same class are classified into one class, so that the semantic segmentation is to understand the image from the pixel level. The high-voltage line detection technology based on the deep learning algorithm is suitable for a high-voltage line bending scene, because each pixel point of an image is classified, the high-voltage line detection technology consists of a group of pixels no matter whether the high-voltage line is bent or straight, and the high-voltage line detection function can be realized as long as the group of pixels are correctly classified.
For the high-voltage line detection method based on the edge attention mechanism fusion network, which is disclosed by the invention, the specific steps are as follows:
1) Network reads the locally stored parameter weights
After the network is subjected to supervision training, parameters are fixed, parameter weight model data are stored locally, and when high-voltage line detection is needed, the network reads the parameter weights from the local parameter weight model data again and loads the parameter weights into the network;
2) Inputting a high-voltage line image to be detected
Inputting a high-voltage line image to be detected into a network, and automatically performing processing operation shown in fig. 1 on the image by the network, wherein the processing operation of the network specifically comprises: dividing the image into two processing paths, extracting characteristic information of the image by using a convolution module, and then carrying out fusion operation on the extracted deep semantic information and shallow space position information by using a characteristic fusion module;
3) Outputting a binary predictive label graph
The convolution module is used for extracting the characteristic information of the image, and the size of the image is reduced, so that after the characteristic fusion module is used for processing, the characteristic image is subjected to bilinear interpolation operation to restore to the original image size, then two-classification operation is performed, whether each pixel of the image belongs to a high-voltage line is judged, finally, a binary predictive label image is output, whether the high-voltage line exists can be judged according to the finally output predictive label image, as shown by d in fig. 3, black pixels represent network prediction as irrelevant background, red pixels represent network prediction as high-voltage line, and the red pixels described herein represent white parts in the image due to the submitted black-white image.
For an input image to be detected, the specific processing procedures of the two processing paths are as follows:
The upper processing path is to perform three continuous downsampling convolution operations on an input image to be detected, each convolution module can enable the size of the input image to be half of the original size, after the three downsampling convolution operations, the size of an output feature image is 1/8 of the size of the original input image, and the path is used for retaining shallow space position information of the image;
The lower processing path is to perform five continuous downsampling convolution operations on the input image to be detected, and similarly, the size of the output feature image is 1/32 of that of the original input image, and the path is used for increasing the receptive field of the network, so that the network obtains higher-level semantic information, and the higher-level semantic information represents deep semantic information.
Then, the output feature graphs of the upper and lower processing paths are input into a feature fusion module, so that the network can better reserve the spatial position information of the image while obtaining higher-level semantic information, and the feature fusion module uses a pixel-level fusion strategy to carry out fusion processing, and the specific operation is as follows: and (3) downsampling the 1/8 feature map output by the upper convolution module into a 1/32 feature map, performing Sigmoid activation on the 1/32 feature map output by the lower convolution module, and finally performing pixel-level multiplication operation with the 1/32 feature map above.
After fusion processing, the output of the feature fusion module is divided into two paths, the lower output path is to restore the 1/32 feature map to the original input image size by continuous up-sampling operation, and then to perform two-classification operation (judging whether each pixel of the image belongs to a high-voltage line or not), namely a predicted main body label output by the model, and calculating a loss value with a real high-voltage line main body label to obtain main body loss;
The upper output path is the key point of the patent, the upper output path through three downsampling convolution modules is used for carrying out resampling operation on the output characteristic image f 1 through an edge information extraction fusion module to obtain a resampled characteristic image f 2, then the output characteristic image f 1 and the resampled characteristic image f 2 are subjected to pixel subtraction operation f 1-f2 to extract a characteristic image f 3 containing edge information, and then the characteristic image f 4 is subjected to edge information fusion with the shallow characteristic image f 1 output by the upper convolution module, so that the space information of f 1 is reserved, the edge information of f 3 is fused, then the f 4 is subjected to upsampling operation to the size of an original input image, and finally two classification operations (judging whether each pixel of the image belongs to a high-voltage line edge) are carried out, namely a predicted edge label output by the model, and a loss value is calculated with a true high-voltage line edge label to obtain edge loss;
for the edge information extraction fusion module, it is based on a policy: image = body + edge. By performing downsampling operation and upsampling operation on the feature map, edge information of the feature map is blurred due to characteristics of the sampling operation, at the moment, main body information of the feature map can be obtained by resampling the blurred feature map, and then the main body information of the feature map is subtracted from the original feature map, so that the edge information of the feature map can be obtained, and the steps are shown in fig. 2. The principle of the fusion module is similar to that of the feature fusion module, and will not be described in detail herein.
And finally, adding the main loss and the edge loss to obtain total loss, and then continuously reducing the total loss by using a gradient descent algorithm, so that the training and learning process of the network is performed, the high-voltage line position information in the image is accurately detected, on the other hand, the parameter weight is continuously adjusted in the training and learning process of the grid, finally, the fixed and invariable optimized parameter is obtained and can be kept locally, and the network can read the parameter weight from the local parameter weight model data again in the next detection process and is loaded in the network.
The main body prediction tag and the edge prediction tag both use Sigmoid activation functions, which are the most commonly used prediction functions in two kinds of networks, and can map all real numbers onto a (0, 1) interval, and the definition formula is as follows:
In formula (1), e is a base of natural logarithms, which is an infinite non-cyclic fraction of 2.71828. z represents a real value obtained by calculation for image input, after the real value is mapped onto a (0, 1) interval through a Sigmoid activation function, whether the mapped value is larger than 0.5 or not can be judged, if the mapped value is larger than 0.5, the network can be predicted to be 1, namely a high-voltage line, and if the mapped value is smaller than or equal to 0.5, the mapped value is predicted to be 0, namely the background.
The body loss and the edge loss both use a cross entropy loss function, which is a function of a calculated loss value most commonly used in deep learning, and the definition formula is as follows:
In the formula (2), N represents the total number of training samples; m represents the number of sample classes; y ic is an illustrative variable, which indicates that when the true category of the category c and the sample i is the same, 1 is set, otherwise, 0 is set; p ic represents the predicted probability that the training sample i belongs to category c.
Because the high-voltage line detection method belongs to the two classification cases, the definition of the cross entropy can be simplified as follows:
y i represents the label of the training sample i, the positive class is 1, and the negative class is 0; p i represents the probability that training sample i is predicted positive.
The high-voltage line detection method based on the edge attention mechanism fusion network is provided for aiming at the technical defects of the traditional high-voltage line detection technology;
Firstly, the basic idea of a high-voltage line detection technology based on a traditional image processing algorithm is divided into three steps, edge information is extracted from an image, pseudo high-voltage line edge information is removed according to priori knowledge of the high-voltage line, and the high-voltage line edge information is fitted into a complete high-voltage line by using a straight line detection technology (Hough transformation or radon transformation). The existing high-voltage line detection technology realized by the traditional image processing algorithm generally comprises the three steps, and the improved thought is to choose to improve the realization effect of one of the steps, such as using a better edge extraction filter to extract edge information, removing more false high-voltage line edge information according to statistical knowledge, and the like. As shown in the steps, the high-voltage line detection technology realized based on the traditional image processing algorithm is complex in steps, more in manual intervention, narrow in applicable scene range and not applicable to high-voltage line bending scenes.
Secondly, the basic idea of the high-voltage line detection technology based on the machine learning algorithm is divided into four steps, the input image is preprocessed, the characteristic information is extracted after the preprocessing, the optimal characteristic information is screened, and the screened characteristic information is input into a machine learning model for training, learning and classifying, so that whether the input image contains the high-voltage line is judged. The quality of the feature information can directly influence the classification performance of the final machine learning model, but as shown in the steps, the steps of extracting the feature information and screening the optimal feature information are all needed to be completed manually, the selected feature information is probably not the optimal feature information, the range of a scene to which the manually selected feature information is applicable is narrow, and the processing steps are complicated.
The invention designs a high-voltage line detection technology based on a deep learning algorithm, and the basic thought is as follows: and designing a proper deep learning model and a loss function, calculating a loss value according to a predicted label and a real label output by the deep learning model for the input image of the deep learning model, and continuously reducing the loss value through a gradient descent algorithm so that the predicted label output by the deep learning model is similar to the real label as much as possible. As shown in the step, the high-voltage line detection technology realized based on the deep learning algorithm does not need to manually select characteristic information, and the optimal characteristic information is automatically learned by the deep learning model, so that the high-voltage line detection technology belongs to an end-to-end model, has simple processing steps and is suitable for a scene of bending of the high-voltage line.
However, the high-voltage line detection technology realized based on the deep learning algorithm also has some problems, such as the problem of inaccuracy of the edge detection of the high-voltage line, and the like, which is determined by the characteristics of the deep learning model, so that in order to alleviate the problem of inaccuracy of the edge detection, the invention adopts the high-voltage line detection method based on the edge attention mechanism fusion network, and can effectively alleviate the problem of inaccuracy of the edge detection of the high-voltage line of the deep learning model.
The high-voltage line detection method based on the edge attention mechanism fusion network is further explained below with reference to specific embodiments.
As shown in the result comparison chart of fig. 3, column (a) represents the original image input to the network, that is, the RGB image required to detect high voltage. (b) The column is a binary real label, red (namely a white part in the figure) corresponds to the position of the high-voltage line, black corresponds to the background, and the binary real label is used for training a network and evaluating the quality of the predictive label through manual labeling. (c) The columns are prediction labels of the semantic segmentation network without the addition of the edge attention mechanism fusion module, red (namely, white parts in the figure) corresponds to the positions of the high-voltage lines, black corresponds to the background, and the problems of fracture and dislocation of the prediction labels can be seen. (d) The list is the prediction label based on the edge attention mechanism fusion network, and compared with the labels predicted by (c) and (d), the prediction label solves the problem of breakage and dislocation well, and the prediction position is accurate.
The invention provides an edge information extraction fusion module based on an edge information extraction process of a feature map for a network, which can extract the edge information of the feature map, and correspondingly the edge information of a high-voltage line to be detected, and the edge information extraction fusion module has the function similar to that of the edge extraction step of the traditional image processing technology. In addition, aiming at the existing deep learning semantic segmentation network model, the edge information extraction fusion module is added in the network to form the fusion network based on the edge attention mechanism, and the edge loss and the main body loss are utilized to train and optimize network parameters at the same time, so that the problem that the existing semantic segmentation network model is inaccurate in detecting the edge of the high-voltage line is greatly solved. Compared with the prior art, the invention has the following advantages:
1) The problem of inaccuracy in detecting the edge of a high-voltage line is solved
Based on the edge attention mechanism fusion network, edge loss is also taken into consideration to train and optimize network parameters, so that the problem that the existing semantic segmentation network model is inaccurate in detecting the edge of the high-voltage line is greatly solved.
2) Simplifying the processing steps
The high-voltage line detection method based on the deep learning algorithm belongs to an end-to-end training model, and compared with the traditional image processing technology and machine learning technology, the high-voltage line detection method based on the deep learning algorithm greatly simplifies processing steps and reduces manual intervention.
3) Be applicable to crooked high-voltage line detection scene
The semantic segmentation model realized based on the deep learning algorithm classifies each pixel of the image, does not adopt any straight line detection technology, and can be suitable for bending high-voltage line detection scenes, which cannot be dealt with by the traditional image processing technology.
The above embodiments are not intended to limit the present invention, and the present invention is not limited to the above examples, but is also intended to be limited to the following claims.
Claims (3)
1. A high-voltage line detection method based on an edge attention mechanism fusion network is characterized by comprising the following steps of: the high-voltage line detection method comprises the following steps:
1) Network reads the locally stored parameter weights
After the network is subjected to supervision training, parameters are fixed, parameter weight model data are stored locally, and when high-voltage line detection is needed, the network reads the parameter weights from the local parameter weight model data again and loads the parameter weights into the network;
2) Inputting a high-voltage line image to be detected
Inputting a high-voltage line image to be detected to a network, and automatically processing the image by the network, wherein the processing operation of the network is specifically as follows: dividing the image into two processing paths, namely an upper processing path and a lower processing path, extracting characteristic information of the image by using a convolution module, and then carrying out fusion operation on the extracted deep semantic information and shallow space position information by using a characteristic fusion module;
3) Outputting a binary predictive label graph
After the feature fusion module is used for processing, performing bilinear interpolation operation on the feature map to restore the original map size, performing two-classification operation, judging whether each pixel of the image belongs to a high-voltage line, and finally outputting a binary predictive label map;
The method comprises the following specific operations that an output feature map of a processing path above and below an image to be detected is input to a feature fusion module, and the feature fusion module uses a pixel-level fusion strategy to carry out fusion processing: the 1/8 feature map output by the upper convolution module is downsampled to become a 1/32 feature map, then the 1/32 feature map output by the lower convolution module is subjected to Sigmoid activation, and finally the 1/32 feature map is multiplied by a pixel level;
The output of the feature fusion module is also divided into two paths, specifically: the lower output path is to restore the 1/32 feature map to the original input image size by continuous up-sampling operation, then to perform two-classification operation to obtain the predicted main body label output by the model, and calculate the loss value with the real high-voltage line main body label to obtain the main body loss;
The upper output path carries out processing operation through an edge information extraction and fusion module, the upper output path through three downsampling convolution modules carries out resampling operation on the output characteristic image, a resampling characteristic image is obtained, then subtraction operation is carried out on the input characteristic image and the resampling characteristic image, an edge information characteristic image is extracted, and edge information fusion is carried out on the input characteristic image and a shallow layer characteristic image output by the convolution module; then up-sampling operation is carried out to restore the original input image size, and finally classification operation is carried out to obtain a predicted edge label output by the model, and a loss value is calculated with the real high-voltage line edge label to obtain edge loss; wherein the shallow feature map is a feature map of the upper processing path output.
2. The high-voltage line detection method based on the edge attention mechanism fusion network according to claim 1, wherein the method comprises the following steps: for an input image to be detected, the specific processing procedures of the two processing paths are as follows:
The upper processing path is to perform three continuous downsampling convolution operations on an input image to be detected, each convolution module can enable the size of the input image to be half of the original size, after the three downsampling convolution operations, the size of an output feature image is 1/8 of the size of the original input image, and the path is used for retaining shallow space position information of the image;
The lower processing path is to perform five continuous downsampling convolution operations on the input image to be detected, and similarly, the size of the output feature image is 1/32 of that of the original input image, and the path is used for increasing the receptive field of the network, so that the network obtains higher-level semantic information, and the higher-level semantic information represents deep semantic information.
3. The high-voltage line detection method based on the edge attention mechanism fusion network according to claim 2, wherein the method comprises the following steps: after the main body loss and the edge loss are obtained, the main body loss and the edge loss are added to obtain total loss, and then the gradient descent algorithm is used for continuously reducing the total loss, so that the training process of the network is carried out, and the high-voltage line position information in the image is accurately detected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011210495.2A CN112329610B (en) | 2020-11-03 | 2020-11-03 | High-voltage line detection method based on edge attention mechanism fusion network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011210495.2A CN112329610B (en) | 2020-11-03 | 2020-11-03 | High-voltage line detection method based on edge attention mechanism fusion network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112329610A CN112329610A (en) | 2021-02-05 |
| CN112329610B true CN112329610B (en) | 2024-07-12 |
Family
ID=74323015
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011210495.2A Active CN112329610B (en) | 2020-11-03 | 2020-11-03 | High-voltage line detection method based on edge attention mechanism fusion network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112329610B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114445436A (en) * | 2021-12-28 | 2022-05-06 | 厦门市美亚柏科信息股份有限公司 | Target detection method, device and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110335270A (en) * | 2019-07-09 | 2019-10-15 | 华北电力大学（保定） | Transmission line of electricity defect inspection method based on the study of hierarchical regions Fusion Features |
| CN110544217A (en) * | 2019-08-30 | 2019-12-06 | 深圳市商汤科技有限公司 | image processing method and device, electronic equipment and storage medium |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB2580671B (en) * | 2019-01-22 | 2022-05-04 | Toshiba Kk | A computer vision system and method |
| CN110930409B (en) * | 2019-10-18 | 2022-10-14 | 电子科技大学 | Salt body semantic segmentation method and semantic segmentation system based on deep learning |
| CN111626298B (en) * | 2020-04-17 | 2023-08-18 | 中国科学院声学研究所 | Real-time image semantic segmentation device and segmentation method |
| CN111666988A (en) * | 2020-05-22 | 2020-09-15 | 哈尔滨理工大学 | Target detection algorithm based on multi-layer information fusion |
-
2020
- 2020-11-03 CN CN202011210495.2A patent/CN112329610B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110335270A (en) * | 2019-07-09 | 2019-10-15 | 华北电力大学（保定） | Transmission line of electricity defect inspection method based on the study of hierarchical regions Fusion Features |
| CN110544217A (en) * | 2019-08-30 | 2019-12-06 | 深圳市商汤科技有限公司 | image processing method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112329610A (en) | 2021-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110276765B (en) | Image panorama segmentation method based on multitask learning deep neural network | |
| CN110969130B (en) | Driver dangerous action identification method and system based on YOLOV3 | |
| CN114841972A (en) | Power transmission line defect identification method based on saliency map and semantic embedded feature pyramid | |
| CN114118124A (en) | Image detection method and device | |
| CN110490155B (en) | Method for detecting unmanned aerial vehicle in no-fly airspace | |
| CN113920468A (en) | Multi-branch pedestrian detection method based on cross-scale feature enhancement | |
| CN111062383A (en) | Image-based ship detection depth neural network algorithm | |
| CN112101153A (en) | Remote sensing target detection method based on receptive field module and multiple characteristic pyramid | |
| CN110991447A (en) | Train number accurate positioning and identification method based on deep learning | |
| CN113128476A (en) | Low-power consumption real-time helmet detection method based on computer vision target detection | |
| CN115937659A (en) | Mask-RCNN-based multi-target detection method in indoor complex environment | |
| CN112329610B (en) | High-voltage line detection method based on edge attention mechanism fusion network | |
| Saravanarajan et al. | Improving semantic segmentation under hazy weather for autonomous vehicles using explainable artificial intelligence and adaptive dehazing approach | |
| CN115527096A (en) | Small target detection method based on improved YOLOv5 | |
| CN113269119B (en) | Night vehicle detection method and device | |
| CN111898608B (en) | Natural scene multi-language character detection method based on boundary prediction | |
| CN117671647B (en) | Multitasking road scene perception method | |
| CN117495825A (en) | Method for detecting foreign matters on tower pole of transformer substation | |
| CN117789077A (en) | Method for predicting people and vehicles for video structuring in general scene | |
| CN113269171B (en) | Lane line detection method, electronic device and vehicle | |
| Zhang et al. | Contextual and Multi-Scale Feature Fusion Network for Traffic Sign Detection | |
| CN116863227A (en) | Hazardous chemical vehicle detection method based on improved YOLOv5 | |
| Jabeen et al. | Weather classification on roads for drivers assistance using deep transferred features | |
| CN112115977B (en) | Target detection algorithm based on scale invariance and feature fusion | |
| CN116434170A (en) | Automatic driving scene vehicle detection method based on improved YOLOv5 |
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 |