An image recognition method and system for the safe construction of power transmission projects

Abstract

The present invention proposes an image recognition method and system for the safe construction of transmission projects. The method includes obtaining image data of fire and smoke hazards in the transmission scene and dividing the image data of fire and smoke hazards into a training set and a test set according to a preset ratio; constructing a detection network YOLOV5 for image recognition; adding a transformer module after the SPPF structure of the network and adding skip connections in the feature fusion layer; preprocessing the training set data and inputting it into the detection network to complete the training of the detection model for the image data of fire and smoke hazards; then testing the test set data, and obtaining the image recognition result for the safe construction of transmission projects after the test. Based on this method, an image recognition system for the safe construction of transmission projects is also proposed. The present invention modifies the detection network structure and the connection method of feature fusion, strengthens the dependence relationship between the global and local parts of the image, enhances the feature extraction ability, and reduces background interference.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent construction of transmission projects, and particularly relates to an image recognition method and system for the safe construction of transmission projects. Background Art

[0002] With the upgrade of transmission line construction technology, visual remote inspection of transmission line corridors has been widely applied. However, due to some existing safety hazards during construction, such as fireworks, etc., the safety of transmission projects is affected, and the progress of transmission project items cannot be guaranteed with high quality. The existing prevention methods for engineering safety hazards still mainly rely on manual inspection, but the labor intensity of manual work is relatively large, and it is impossible to effectively prevent hazards throughout the day. Therefore, a new type of image recognition technology is needed to detect the surrounding environment of transmission lines in real time, thereby preventing safety hazards, ensuring the safety and quality during the construction process of transmission projects, and enabling safe and civilized construction.

[0003] In recent years, with the increasing improvement of computer and GPU computing capabilities, deep learning has been successfully applied in a variety of computer vision tasks. Among tens of thousands of image recognition tasks, deep learning has a fast recognition speed and far better recognition accuracy than humans. How to provide an efficient and reliable transmission image recognition technology that can quickly and effectively ensure the safety and high quality of transmission projects and prevent safety hazards is one of the problems that need to be solved urgently by technical personnel in this field at present. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes an image recognition method and system for the safe construction of transmission projects. It realizes the detection of safety hazards during the construction process of transmission projects and returns the detection results to the monitoring personnel, reducing the workload of the monitoring personnel and eliminating the hazards during the construction process of transmission projects.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An image recognition method for the safe construction of transmission projects includes the following steps:

[0007] Obtain the image data of fireworks hazards in the transmission scene and divide the image data of fireworks hazards into a training set and a test set according to a preset ratio;

[0008] Construct the detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure and adds skip connections in the feature fusion layer;

[0009] Preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fireworks hazards;

[0010] Use the trained detection model to test the test set data, and obtain the image recognition result of the safe construction of the transmission project after testing.

[0011] Furthermore, the ratio of the training set to the test set is 8:2.

[0012] Furthermore, the process of the transformer module working includes: converting the feature map after target feature extraction into the form of input-output sequences, and then performing optimization calculations through the multi-head self-attention layer to linearly connect multiple attention outputs to the desired dimension.

[0013] Furthermore, the process of the preprocessing is: scaling the image data of the fire hazard to the size required by the detection network, and performing normalization processing.

[0014] Furthermore, the size required by the detection network is 1280*1280.

[0015] Furthermore, after the preprocessing, it also includes:

[0016] Use Mosaic for data augmentation, and splice in a random scaling, random cropping or random arrangement manner to increase the generalization of the model, and perform adaptive anchor box calculation;

[0017] Use K-means clustering to update the anchor to adapt to targets of different sizes in the same image.

[0018] Furthermore, after using K-means clustering to update the anchor, it also includes: using the genetic algorithm to re-search for training hyperparameters, re-adjusting the network parameters for training, and using the adam optimizer for gradient update.

[0019] Furthermore, the process of training the detection model for the image data of the fire hazard includes:

[0020] First, perform feature extraction of the image through the Backbone stage. After a series of convolutions and activations, the training set image data undergoes downsampling to obtain the feature map;

[0021] Secondly, learn the correlation between features through the transformer; repeat upsampling multiple times in the Neck stage, and perform multiple skip connections to strengthen the fusion of feature information and make up for the loss of detailed information caused by downsampling;

[0022] Finally, perform high-level semantic information extraction in the Output stage, and regress the target area and category information at different scales.

[0023] The present invention also provides an image recognition system for the safe construction of a power transmission project, which includes an acquisition module, a construction module, a training module, and a testing module;

[0024] The acquisition module is used to acquire image data of fire and smoke hazards in a power transmission scenario, and divide the image data of fire and smoke hazards into a training set and a testing set according to a preset ratio;

[0025] The construction module is used to construct a detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure, and adds skip connections in the feature fusion layer;

[0026] The training module is used to preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards;

[0027] The testing module is used to test the testing set data with the trained detection model, and obtain the image recognition result of the safe construction of the power transmission project after testing.

[0028] Further, the process of the transformer module working includes: converting the feature map after target feature extraction into the form of an input-output sequence, and then performing optimization calculation through a multi-head self-attention layer to linearly connect multiple attention outputs to the desired dimension.

[0029] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0030] The present invention provides an image recognition method and system for the safe construction of a power transmission project. The method includes acquiring image data of fire and smoke hazards in a power transmission scenario, and dividing the image data of fire and smoke hazards into a training set and a testing set according to a preset ratio; constructing a detection network YOLOV5 for image recognition; adding a transformer module after the SPPF structure in the detection network YOLOV5, and adding skip connections in the feature fusion layer; preprocessing the training set data and inputting it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards; testing the testing set data with the trained detection model, and obtaining the image recognition result of the safe construction of the power transmission project after testing. Based on an image recognition method for the safe construction of a power transmission project, an image recognition system for the safe construction of a power transmission project is also provided. The present invention combines the transformer module with the existing detection network, modifies the connection method of feature fusion, strengthens the global and local dependence relationship of the image, enhances the feature extraction ability, and reduces background interference.

[0031] The present invention improves the robustness of the overall detection algorithm, has better foreground extraction ability in complex scenarios, and also has good detection effect on small targets.

[0032] The present invention can efficiently and reliably detect potential safety hazards in the monitoring area, providing technical support and guarantee for the safety during the construction of transmission projects and the high-quality promotion of project items. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] As Figure 1 is a flowchart of an image recognition method for the safety construction of transmission projects in Embodiment 1 of the present invention;

[0034] As Figure 2 is an architecture diagram of the detection network YOLOV5 for image recognition in Embodiment 1 of the present invention;

[0035] As Figure 3 is a schematic diagram of an image recognition system for the safety construction of transmission projects in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To clearly illustrate the technical features of this solution, the present invention will be elaborated in detail below through specific embodiments and in conjunction with its accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. The present invention omits the description of well-known components and processing technologies and processes to avoid unnecessarily limiting the present invention.

[0037] Embodiment 1

[0038] Embodiment 1 of the present invention proposes an image recognition method for the safety construction of transmission projects, which uses deep learning algorithms to control the construction of transmission infrastructure projects, realizes the detection of potential safety hazards during the construction of transmission projects, and returns the detection results to the monitoring personnel, reducing the workload of the monitoring personnel and eliminating potential hazards during the construction of transmission projects. The present invention can efficiently and reliably detect abnormal potential safety hazards and is well applicable to the control process of transmission infrastructure projects.

[0039] As Figure 1 is a flowchart of an image recognition method for the safety construction of transmission projects in Embodiment 1 of the present invention;

[0040] In step S100, obtain the image data of fire and smoke hazards in the power transmission scenario, and divide the image data of fire and smoke hazards into a training set and a test set according to a preset ratio; the ratio of the training set to the test set is 8:2.

[0041] In step S110, construct the detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure, and adds skip connections in the feature fusion layer;

[0042] As Figure 2 is the architecture diagram of the detection network YOLOV5 for image recognition in Embodiment 1 of the present invention. A transformer module is added after the SPPF structure, the feature fusion layer is modified, two skip connections are added, the flow of feature information in the context is increased, and the robustness of the network is improved.

[0043] First, use CNN to extract significant features such as the texture and color of the target, and then convert the feature map into a processing method similar to tokens, so that it can be in the form of sequences for both input and output. Through self-attention parallel computing, the dependency relationship between input and output is mined. The multi-head self-attention layer can linearly connect multiple attention outputs to the desired dimension, which helps to understand the local and global dependency relationships of the image. Later, in order to enhance the learning ability of CNN, feature extraction is performed on the branch path, the feature map is convolved, and the results before and after convolution are concatenated.

[0044] The processing method similar to tokens is as follows: perform data transformation through the Embedding layer. For a feature map of H*W, divide it into patches of S*S, and then obtain K=(H / S)*(W / S) patches. Then map the patches into one-dimensional vectors through linear mapping, and the vector with a length of S*S obtained by mapping is the token.

[0045] In order to increase the fusion of feature information between upper and lower layers, add a skip connection to the input and output nodes at different scales. While not increasing too much computational cost, more features are fused. Each bidirectional path is regarded as a feature network layer, and the same layer is repeated multiple times to achieve high-level feature fusion.

[0046] To solve the problem that the IOU loss has a gradient of 0 when the predicted box and the ground truth box do not overlap, use GIOU loss, which can well reflect the overlap degree of the two.

[0047] In step S120, preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards;

[0048] The image is preprocessed, scaled to the required size of 1280*1280 for the detection network, and normalized to reduce the amount of calculation, enabling the model to converge quickly;

[0049] Mosaic is used for data augmentation, and splicing is performed in a way of random scaling, random cropping, and random arrangement, thereby increasing the generalization of the model, and adaptive anchor box calculation is carried out. K-means clustering is used to update the anchor to adapt to targets of different sizes in the same image.

[0050] The genetic algorithm is used to re-search for training hyperparameters, the network parameters are re-adjusted for training, and the adam optimizer is used for gradient update.

[0051] First, image feature extraction is performed through the Backbone stage. After a series of convolutions and activations, the training set image data is downsampled to obtain a feature map; second, the transformer is used to learn the correlation between features; in the Neck stage, upsampling is repeated multiple times and multiple skip connections are made to strengthen the fusion of feature information and make up for the loss of detailed information caused by downsampling; finally, in the Output stage, high-level semantic information is extracted, and the target region and class information are regressed at different scales.

[0052] General features of the image can be quickly learned in the initial stage of training. In the later stage, the upper-layer weight parameters can be selected to be frozen, and the lower-layer weight values can be fine-tuned so that the model can focus more on the areas to be recognized.

[0053] In step S130, the trained detection model is used to test the test set data, and after testing, the image recognition result of the transmission project safety construction is obtained.

[0054] An image recognition method for transmission project safety construction proposed in Embodiment 1 of the present invention combines the transformer module with the existing detection network, modifies the connection method of feature fusion, strengthens the global and local dependence relationships of the image, enhances the feature extraction ability, and reduces background interference.

[0055] An image recognition method for transmission project safety construction proposed in Embodiment 1 of the present invention improves the robustness of the overall detection algorithm, has better foreground extraction ability in complex scenarios, and also has good detection effect on small targets.

[0056] An image recognition method for transmission project safety construction proposed in Embodiment 1 of the present invention can efficiently and reliably detect potential safety hazards in the monitoring area, providing technical support and guarantee for the safety during the transmission project construction process and the high-quality promotion of the engineering project.

[0057] Embodiment 2

[0058] Based on the image recognition method for the safe construction of transmission projects proposed in Embodiment 1 of the present invention, Embodiment 2 of the present invention also proposes an image recognition system for the safe construction of transmission projects. As Figure 3 FIG. is a schematic diagram of an image recognition system for the safe construction of transmission projects according to Embodiment 2 of the present invention. The system includes an acquisition module, a construction module, a training module, and a testing module;

[0059] The acquisition module is used to acquire image data of fire and smoke hazards in the transmission scenario and divide the image data of fire and smoke hazards into a training set and a testing set according to a preset ratio; the ratio of the training set to the testing set is 8:2.

[0060] The construction module is used to construct the detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure and adds skip connections in the feature fusion layer;

[0061] The training module is used to preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards;

[0062] The testing module is used to test the testing set data with the trained detection model, and obtain the image recognition result of the safe construction of the transmission project after testing.

[0063] The process implemented by the construction module includes: As Figure 2 FIG. is the architecture diagram of the detection network YOLOV5 for image recognition according to Embodiment 1 of the present invention. A transformer module is added after the SPPF structure, and the feature fusion layer is modified to add two skip connections to increase the flow of feature information in the context and improve the robustness of the network.

[0064] First, use CNN to extract significant features such as the texture and color of the target, and then convert the feature map into a processing method similar to tokens, so that it can be in the form of sequences for both input and output. Through self-attention parallel computing, the dependency relationship between input and output is mined. The multi-head self-attention layer can linearly connect multiple attention outputs to the desired dimension, which helps to understand the local and global dependency relationships of the image. Later, in order to enhance the learning ability of CNN, feature extraction is performed on the branch path, the feature map is convolved, and the results before and after convolution are concatenated.

[0065] The processing method of similar tokens is as follows: data transformation is performed through the Embedding layer. For a feature map of H*W, it is divided into patches of S*S, and then K = (H / S)*(W / S) patches are obtained. Then, the patches are linearly mapped into one-dimensional vectors, and the vector with a length of S*S obtained by the mapping is the token.

[0066] To increase the fusion of feature information between the upper and lower layers, a skip connection is added to the input and output nodes at different scales. While not increasing too much computational cost, more features are fused. Each bidirectional path is regarded as a feature network layer, and the same layer is repeated multiple times to achieve high-level feature fusion.

[0067] To solve the problem that the loss gradient of the IOU loss is 0 when the predicted box and the ground truth box do not overlap, the GIOU loss is adopted, which can well reflect the coincidence degree of the two.

[0068] The process implemented by the training module includes: preprocessing the image, scaling it to the required size of 1280*1280 for the detection network, and normalizing it to reduce the computational amount and enable the model to converge quickly;

[0069] Data augmentation is performed using Mosaic, and splicing is carried out in a way of random scaling, random cropping, and random arrangement, so as to increase the generalization of the model, and adaptive anchor box calculation is performed. The K-means clustering is used to update the anchor to adapt to targets of different sizes in the same image.

[0070] The genetic algorithm is used to re-search for training hyperparameters, the network parameters are re-adjusted for training, and the adam optimizer is used for gradient update.

[0071] First, image feature extraction is performed through the Backbone stage. After a series of convolutions and activations, the training set image data is downsampled to obtain a feature map. Secondly, the correlation between features is learned through the transformer. In the Neck stage, upsampling is repeated multiple times and multiple skip connections are made to strengthen the fusion of feature information and make up for the loss of detailed information caused by downsampling. Finally, in the Output stage, high-level semantic information is extracted, and the target region and class information are regressed at different scales.

[0072] In the initial stage of training, the general features of the image can be quickly learned. In the later stage, the upper layer weight parameters can be frozen and the lower layer weight values can be fine-tuned, so that the model can focus more on the regions to be recognized.

[0073] An image recognition system for the safe construction of power transmission projects proposed in Embodiment 2 of the present invention combines the transformer module with the existing detection network, modifies the connection method of feature fusion, strengthens the global and local dependence relationships of the image, enhances the feature extraction ability, and reduces background interference.

[0074] An image recognition system for the safe construction of power transmission projects proposed in Embodiment 2 of the present invention improves the robustness of the overall detection algorithm, has better foreground extraction ability in complex scenarios, and also has good detection effect on small targets.

[0075] An image recognition system for the safe construction of power transmission projects proposed in Embodiment 2 of the present invention can efficiently and reliably detect potential safety hazards in the monitoring area, providing technical support and guarantee for the safety during the construction of power transmission projects and the high-quality promotion of engineering projects.

[0076] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes the inherent elements thereof. Without more limitations, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element. In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the corresponding technical solutions in the prior art in terms of implementation principles are not described in detail to avoid excessive elaboration.

[0077] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. For those skilled in the art, other different forms of modification or variation can be made on the basis of the above description. It is not necessary and impossible to enumerate all the implementation manners here. Based on the technical solutions of the present invention, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. An image recognition method for the safe construction of transmission projects, characterized in that, It includes the following steps: Obtain the image data of fire and smoke hazards in the power transmission scenario, and divide the image data of fire and smoke hazards into a training set and a test set according to a preset ratio; Construct the detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure and adds skip connections in the feature fusion layer; the process of the transformer module working includes: converting the feature map after target feature extraction into the form of input and output sequences, and then performing optimization calculations through the multi-head self-attention layer to linearly connect multiple attention outputs to the desired dimension; adding two skip connections to increase the flow of feature information of the context; Preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards; the process of training the detection model for the image data of fire and smoke hazards includes: first, perform feature extraction of the image through the Backbone stage, and after a series of convolutions and activations, the training set image data is downsampled to obtain a feature map; second, learn the correlation between features through the transformer; in the Neck stage, perform upsampling multiple times and perform multiple skip connections to strengthen the fusion of feature information and make up for the loss of detailed information caused by downsampling; finally, in the Output stage, extract high-level semantic information and regress the target area and category information at different scales; Use the trained detection model to test the test set data, and obtain the image recognition result of the safety construction of the power transmission project after testing.

2. The image recognition method for safe construction of a power transmission project according to claim 1, wherein, The ratio of the training set to the test set is 8:

2.

3. The image recognition method for safe construction of a power transmission project according to claim 1, characterized in that The process of the preprocessing is: scale the image data of fire and smoke hazards to the size required by the detection network and perform normalization processing.

4. The image recognition method for safe construction of power transmission projects according to claim 3, wherein, The size required by the detection network is 1280*1280.

5. The image recognition method for safe construction of a power transmission project according to claim 3, characterized in that, After the preprocessing, it further includes: Use Mosaic for data augmentation, and splice in a random scaling, random cropping or random arrangement manner to increase the generalization of the model, and perform adaptive anchor box calculation; Use K-means clustering to update the anchor to adapt to targets of different sizes in the same image.

6. The image recognition method for safe construction of a power transmission project according to claim 5, characterized in that, After using K-means clustering to update the anchor, it further includes: use the genetic algorithm to re-search the training hyperparameters, re-adjust the network parameters for training, and use the adam optimizer for gradient update.

7. An image recognition system for the safe construction of a power transmission project, which is used to execute an image recognition method for the safe construction of a power transmission project described in any one of claims 1 to 6, characterized in that, It includes an acquisition module, a construction module, a training module and a testing module; The acquisition module is used to obtain the image data of fire and smoke hazards in the power transmission scenario, and divide the image data of fire and smoke hazards into a training set and a test set according to a preset ratio; The construction module is used to construct the detection network YOLOV5 for image recognition; the detection network YOLOV5 adds a transformer module after the SPPF structure and adds skip connections in the feature fusion layer; The training module is used to preprocess the training set data and input it into the detection network YOLOV5 to complete the training of the detection model for the image data of fire and smoke hazards; The test module is used to test the test set data by using the trained detection model, and after testing, the image recognition result of the safe construction of the transmission project is obtained.

Images