Decouple-YOLO-based train track fastener missing identification method

Abstract

The invention provides a train track fastener missing identification method based on Decouple-YOLO (Young Occupied Line Occupiers) and relates to a train track fastener missing identification method. The method comprises the following steps: carrying out head detection and decoupling operation on train track fastener image data to be identified to obtain features of a detected head; inputting the feature map of the detected head into a RepVGG backbone network, and outputting image features extracted through the RepVGG backbone network; and inputting the feature map of the detected head and the image features extracted by the RepVGG backbone network into a backbone network based on ShuffleNetv2, and outputting a train track fastener missing identification result by the backbone network based on the ShuffleNetv2. Decoupling operation is performed on the detection head for an unmanned aerial vehicle inspection application scene to solve conflicts between classification and regression tasks, and the detection precision and the detection speed of the model are improved, so that the overall detection performance is improved. According to the invention, the Repblock-C3 module is used as a backbone network, and the detection precision is further improved under the condition that the size and parameter quantity of the model are reduced in cooperation with a decoupling detection head model.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for identifying missing train track fasteners based on Decouple-YOLO. Background Art

[0002] With the increase in train traffic and load, the increase in operating years and the combined effects of various environmental factors, defects will occur in key components of ballastless tracks such as fasteners and track plates, which may lead to traffic accidents in severe cases. Therefore, efficient and accurate inspection of trains is of great significance to ensure safe and stable railway operation.

[0003] If we only rely on drones to collect images of components in the high-speed rail ballastless track area, transmit them to the ground, and then manually analyze the train frame by frame, this is obviously not an efficient and intelligent high-speed rail line infrastructure inspection solution.

[0004] In order to improve the intelligent analysis level of the detection system, how to detect missing fasteners with high precision and high speed is an urgent problem to be solved. Summary of the invention

[0005] The embodiment of the present invention provides a method for identifying missing fasteners based on Decouple-YOLO, so as to effectively identify missing fasteners on train tracks.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme.

[0007] A train track fastener missing identification method based on Decouple-YOLO, comprising:

[0008] Perform head detection and decoupling operations on the train track fastener image data to be identified, and obtain the features of the detected head;

[0009] Inputting the feature map of the detected head into the RepVGG backbone network, and outputting the image features extracted by the RepVGG backbone network;

[0010] The feature map of the detection head and the image features extracted by the RepVGG backbone network are input into the backbone network based on ShuffleNetv2, and the backbone network based on ShuffleNetv2 outputs the train track fastener missing recognition result.

[0011] Preferably, the step of performing head detection and decoupling operations on the train track fastener image data to be identified and obtaining the features of the detected head includes:

[0012] The image data of train track fasteners to be identified is collected by drones, and a 1×1 convolution is used to reduce the dimension of the train track fastener image data. Then, two 3×3 convolutions are used in the classification and regression branches respectively to perform head detection and decoupling operations on the train track fastener image data. Parameters are added on the basis of the original detection head to generate a feature map of the decoupled detection head.

[0013] Preferably, the step of inputting the feature map of the detected head into a RepVGG backbone network and outputting the image features extracted by the RepVGG backbone network comprises:

[0014] The feature map of the detected head is input into the RepVGG backbone network. In the RepVGG backbone network, the 1×1 convolution kernel is converted into a 3×3 convolution kernel by padding. For the identity branch, the current convolution kernel parameter is set to 1, and the other convolution kernel parameters are set to 0. The identity branch is first converted into a 1×1 convolution kernel, and then converted into a 3×3 convolution kernel by the above padding operation. Finally, the 3×3 convolutions of the three output scales are added together by using the additivity of convolution to obtain the final convolution kernel.

[0015] The RepVGG backbone network extracts features from the feature map of the detection head, and adjusts the feature map of the detection head to the size required by the network through standardization. During the training phase, each convolution layer of RepVGG consists of multiple parallel convolution paths. The 3x3 convolution branch is used to extract local spatial features, the 1x1 convolution branch is used to mix feature channels, and the identity mapping branch is used to maintain the residual connection of the features. The outputs of all branches are point-by-point additively fused, and nonlinearity is introduced through the ReLU activation function. The last layer of the RepVGG network outputs compressed image features, which capture high-level semantic information in the input image.

[0016] Preferably, the step of inputting the feature map of the detection head and the image features extracted by the RepVGG backbone network into a backbone network based on ShuffleNetv2, and the backbone network based on ShuffleNetv2 outputting a train track fastener missing recognition result, comprises:

[0017] A backbone network based on ShuffleNetV2 is constructed according to four convolutional network design criteria, which include: (G1) when the number of input feature channels of the convolutional layer is equal to the number of output feature channels, the memory access time MAC is minimized; (G2) using grouped convolution with a large number of groups will increase MAC; (G3) fragmented parallel operations will reduce the model inference speed; (G4) considering the memory usage and time consumption caused by element-by-element operations;

[0018] The feature map of the detected head and the image features extracted by the RepVGG backbone network are input into the backbone network based on ShuffleNetv2. The ShuffleNetv2 network linearly combines the features of different channels through point-by-point convolution, and splices the features from different networks together. The channels of the input feature map are divided into multiple groups, and each group performs a convolution operation separately. ShuffleNetv2 mixes the channels after convolution by group, so that the information between different groups interacts with each other. Each channel performs convolution separately to extract spatial features, and uses 1x1 convolution to combine the features of different channels to achieve channel information fusion. The convolution output is normalized to generate an optimized feature map;

[0019] Subsequent missing fastener detection is performed based on the above-mentioned optimized feature map to obtain a train track fastener missing recognition result, which includes the detected fastener position, the state of the fastener missing, the precision rate and the recall rate information.

[0020] The optimized feature map is input into the detection model, and the detection model decodes the optimized feature map to generate a detection frame and a category probability, and predicts the detection result of the fastener in each detection frame, calculates the confidence of each detection frame according to the category probability of each detection frame, filters out the detection frames with a confidence lower than a set threshold, and uses a non-maximum suppression algorithm to remove redundant detection frames to obtain a train track fastener missing recognition result, which includes the detected fastener position, the status of the fastener missing, the precision rate and the recall rate information.

[0021] It can be seen from the technical solutions provided by the above-mentioned embodiments of the present invention that the present invention decouples the detection head for UAV inspection application scenarios to resolve the conflict between classification and regression tasks, thereby improving the detection accuracy and speed of the model and thus improving the overall detection performance.

[0022] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0024] Figure 1 A processing flow chart of a method for identifying missing train track fasteners based on Decouple-YOLO proposed in an embodiment of the present invention;

[0025] Figure 2 A first backbone network diagram based on RepVGG provided by an embodiment of the present invention;

[0026] Figure 3 A second backbone network diagram based on RepVGG provided in an embodiment of the present invention

[0027] Figure 4 A backbone network diagram based on ShuffleNetv2 provided in an embodiment of the present invention;

[0028] Figure 5 A comparison chart of detection accuracy of different algorithm models provided by an embodiment of the present invention;

[0029] Figure 6 A schematic diagram of a train track fastener missing detection result provided by an embodiment of the present invention;

[0030] Figure 7 A schematic diagram of a precision-recall curve provided by an embodiment of the present invention;

[0031] Figure 8 A heat map of a training process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0033] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0034] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0035] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0036] This paper proposes a method for fastener missing recognition based on deep learning. The decouple-YOLO target detection algorithm based on deep learning is used to decouple the detection head to improve the overall detection performance of fastener defects. The backbone network is designed based on RepVGG and ShuffleNetv2. The model detection performance is improved by optimizing and improving the design of the backbone network and the detection head respectively. The specific steps are as follows:

[0037] The processing flow of a train fastener missing identification method based on Decouple-YOLO proposed in an embodiment of the present invention is as follows: Figure 1 As shown, the processing steps include the following:

[0038] Step S10: performing head detection and decoupling operations on the train track fastener image data to be identified, and obtaining the features of the detected head.

[0039] The original train fastener image data to be identified is collected by drones. The collection of these train fastener image data is usually performed by drones, which cruise over the high-speed railway line and take images of components in the track area. These images are then transmitted to the ground end for further analysis and processing.

[0040] A 1×1 convolution is used to reduce the dimension of the original train fastener image data, and then two 3×3 convolutions are used in the classification and regression branches respectively to perform head detection and decoupling operations on the original train fastener image data. A small number of parameters are added to the original detection head to generate feature maps of the decoupled detection head. These feature maps will be used for subsequent backbone network design and fastener missing identification.

[0041] Step S20: input the feature map of the detected head into the RepVGG backbone network, and output the image features extracted by the RepVGG backbone network.

[0042] The feature map of the above detection head is input to the RepVGG backbone network, and the image features extracted by the RepVGG backbone network are output. These image features will be used for further processing and analysis in subsequent steps to identify the missing train fasteners. The input image is normalized (such as subtracting the mean and dividing by the standard deviation) and adjusted to the size required by the network (224×224). Data augmentation (flipping, scaling, etc.) may be used during the training phase to improve the generalization ability of the model. During the training phase, each convolutional layer of RepVGG consists of multiple parallel convolution paths. 3x3 convolution branch: used to extract local spatial features. 1x1 convolution branch: used to mix feature channels to improve feature expression capabilities. Identity mapping (Identity branch): used to maintain the residual connection of features. The outputs of these branches are point-by-point additively fused, and nonlinearity is introduced through the ReLU activation function. RepVGG uses continuous convolutional layers (3x3 convolutions) to continuously compress the spatial dimensions and increase feature channels to extract higher and higher levels of semantic features. Pooling operations are used between some layers to reduce the spatial resolution of feature maps, reduce the amount of computation, and capture a larger receptive field. The multi-branch structure in the model allows the network to learn richer features. RepVGG fuses the parameters of the multi-branch convolutional layer and reparameterizes it into an equivalent single convolutional layer. This process significantly improves the inference speed while maintaining the accuracy of the model. The last layer of the network outputs a compressed high-dimensional feature map that captures the high-level semantic information in the input image.

[0043] The RepVGG backbone network will have three different scales of output in the inference stage. The present invention converts the 1×1 convolution kernel into a 3×3 convolution kernel through padding. For the identity branch, the current convolution kernel parameter = 1, and the other convolution kernel parameters = 0, and it is first converted into a 1×1 convolution kernel, and then converted into a 3×3 convolution kernel through the above padding operation. Finally, using the additivity of convolution, the 3×3 convolutions of the three output scales are added to obtain the final convolution kernel.

[0044] The present invention designs two types of backbone networks based on RepVGG to extract the features of the input image. The first backbone network directly reuses the RepVGG-A1 model and removes the Focus module at the beginning and the SPPF module at the end to reduce the parallelism of the network and avoid increasing the complexity of the model. Figure 2 The second backbone network uses a stack of C3 modules and RepVGGblock, as shown in Figure 3 The SPPF module at the end of the original backbone network is removed, and the Focus module is replaced by a convolution kernel with a size of 6×6, a step size of 2, and a padding of 2 to improve the portability of the model.

[0045] The C3 module achieves efficient feature fusion and information reuse through multi-path design and residual connection, enabling the model to capture local and global features at the same time and alleviate the gradient vanishing problem. Its stacking with the RepVGG Block enhances the model's feature expression ability, allowing the network to have faster inference speed and lower computational overhead while maintaining detection performance. Removing the SPPF module and replacing the Focus module with a convolution kernel with a step size of 2 and padding of 2 further simplifies the model structure and improves the portability and deployment efficiency of the model on a variety of hardware. This design ensures that the model achieves efficient and accurate target recognition in the task of missing train track fastener detection.

[0046] Step S30: input the feature map of the detection head and the image features extracted by the RepVGG backbone network into the backbone network based on ShuffleNetv2, and the backbone network based on ShuffleNetv2 outputs the train track fastener missing recognition result.

[0047] The feature map of the detection head and the image features extracted by the RepVGG backbone network are input into the backbone network based on ShuffleNetv2. The ShuffleNetv2 backbone network further processes these input data to improve the lightweight and efficiency of the network. In the detection task, the image features extracted by RepVGG and the feature map generated by the detection head are input into the ShuffleNetv2 network as input. These feature maps may come from multi-scale features of different layers, and need to be fused channels first. The specific methods include: point-by-point convolution (1x1 convolution): linearly combine the features of different channels. Concatenation: directly concatenate the features from different networks together. The channels of the input feature map are divided into multiple groups, and each group performs convolution operations separately. ShuffleNetv2 mixes the channels after convolution by group, so that the information between different groups interacts with each other. Each channel performs convolution separately to extract spatial features. And use 1x1 convolution to combine the features of different channels to achieve channel information fusion. In some network layers, ShuffleNetv2 uses convolution or pooling operations with a stride of 2 for downsampling. Nonlinearity is introduced to improve the feature expression ability of the network. The convolution output is normalized to speed up the network training speed and stabilize the network training process. After being processed by the ShuffleNetv2 network, a feature map is generated. Finally, the output data of this step is a further optimized feature map.

[0048] Subsequent missing fastener detection is performed based on the above-mentioned optimized feature map to obtain a train track fastener missing recognition result, which includes information such as the detected fastener position, the state of the fastener missing (such as normal, missing), precision, recall rate, etc.

[0049] Based on the optimized feature map generated by the ShuffleNetv2 backbone network, the system combines the classification and positioning results of the detection head to achieve efficient detection of missing train track fasteners. The optimized feature map is input into the detection model, which decodes the feature map to generate detection boxes and category probability information, and predicts the status of the fasteners in each detection box. In order to improve the detection accuracy, the system filters out detection results below the set threshold based on the predicted confidence, and uses the non-maximum suppression (NMS) algorithm to remove redundant detection boxes to ensure that only the highest confidence results are retained for each fastener position. The detection results include the specific location and status of the fastener, as well as the confidence of each detection result, which are used to analyze the reliability of the model. Through multi-scale feature extraction and information fusion, the system can handle fasteners of different sizes and positions, effectively improving the detection performance of the model. In order to comprehensively evaluate the system performance, the system calculates the precision and recall, which represent the proportion of correctly detected missing fasteners and the proportion of successfully detected missing fasteners, respectively, and uses their harmonic mean (F1-score) to comprehensively evaluate the overall effect of the model. The detection results are presented in a visual form on the track image, with different colored borders marking the fastener status, such as green for normal and red for missing, and structured data is output in JSON or CSV format, recording information such as the detection frame coordinates, status, and confidence. Ultimately, through the analysis of false positives and false negatives, the model can be continuously iterated and optimized to further improve the accuracy and stability of detection, providing reliable data support for the maintenance and management of train track fasteners.

[0050] The backbone network of ShuffleNetV2 is constructed according to four convolutional network design criteria. The above four convolutional network design criteria include: (G1) When the number of input feature channels of the convolution layer is equal to the number of output feature channels, the memory access consumption time (Memory Access Cos, MAC) is minimized; (G2) Using group convolution with a large number of groups will increase MAC; (G3) Fragmented parallel operations (too many network branches) will reduce the model inference speed; (G4) The number of floating-point operations per second (FLOPs) of element-wise operations such as ReLU and shortcut-add is relatively small, but the memory usage and time consumption they bring are not negligible when designing the backbone network.

[0051] The present invention provides a backbone network based on ShuffleNetV2. Figure 4As shown in the figure. First, avoid continuous reuse of the C3 module because its multi-branch structure violates the G3 principle, and in the corresponding operation of the C3 module, the number of feature channels of the input image often changes, which also violates the G1 principle. Secondly, a convolution kernel with a size of 6×6, a step size of 2, and a padding of 2 is used to replace the Focus module, and further remove the SPPF module at the end of its backbone network because this structure contains parallel operations, which violates the G3 principle.

[0052] The present invention compares the Decouple-YOLO-based missing fastener identification method proposed in the present invention and other deep learning algorithms, such as Yolov5s and Yolox, on the fastener data set collected by the drone.

[0053] Figure 5 A comparison chart of detection accuracy of different algorithm models provided by an embodiment of the present invention, such as Figure 5 As shown in the figure, when the backbone network is replaced with a design based on ShuffleNetV2, the accuracy decreases, but when the detection head is decoupled, the accuracy increases. When the backbone network is replaced with a design based on Repblock-C3, the detection accuracy is improved. After the detection head is decoupled, the accuracy is higher than the mAP value of the original YOLOv5s algorithm. Figure 6 A schematic diagram of a train track fastener missing detection result provided by an embodiment of the present invention, such as Figure 6 As shown, both normal fasteners and missing fasteners have good detection effects.

[0054] Figure 7 A schematic diagram of a precision-recall curve provided by an embodiment of the present invention is shown in FIG. Figure 7 As shown in the figure, the PR curve is drawn based on the precision and recall of the normal fastener and missing fastener categories in the test set. The intersection over union (IoU) of the target real box and the predicted box is taken as the correct detection result when the threshold is above 0.5. The average precision of each category is 0.995, which has a good detection result. Figure 8 A training process heat map provided by an embodiment of the present invention is a training process heat map, such as Figure 8 As shown, more attention is paid to the features of the noraml area in the image during training.

[0055] The quantitative evaluation results of the improved model formed by the combination of the backbone network designed based on ShuffleNetV2 and RepVGG and the decoupled detection head and the original YOLOv5 and YOLOX are shown in Table 1. When the backbone network is replaced with a design based on ShuffleNetV2, the size and number of parameters of the model are greatly reduced by 76% and 78%, while the mAP index drops from 74.8 to 72.9. After the detection head is decoupled, the mAP rises back to 74.2, which is 0.6 lower than the mAP value of the original YOLOv5s algorithm, and its model size is only half of that of YOLOv5s. When the backbone network is replaced with a design based on Repblock-C3, the model size and number of parameters are reduced by 21% and 22%, while the mAP index rises from 74.8 to 75.2, and the detection accuracy is improved. After decoupling the detection head, the mAP can reach an accuracy of 75.8, which is 1.0 higher than the mAP value of the original YOLOv5s algorithm. At the same time, the model size and number of parameters are reduced by 14% and 15% respectively compared with the original YOLOv5s model.

[0056] Table 1 Comparison of different model parameters and test results

[0057]

[0058] In summary, the embodiments of the present invention decouple the detection head for drone inspection application scenarios to resolve the conflict between classification and regression tasks, improve the detection accuracy and speed of the model, and thus improve the overall detection performance.

[0059] The present invention uses the Repblock-C3 module as the backbone network and combines it with a decoupled detection head model to further improve detection accuracy while reducing the model size and parameter quantity.

[0060] The size of the network model based on the ShuffleNetV2 backbone design is greatly reduced, and the Focus module and SPPF module are eliminated, which not only reduces parallel operations, reduces memory usage, and improves the detection rate, but also increases the portability of the network model.

[0061] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0062] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0063] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0064] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A train track fastener missing identification method based on Decouple-YOLO, characterized in that: include: Perform head detection and decoupling operations on the train track fastener image data to be identified, and obtain the features of the detected head; Inputting the feature map of the detected head into the RepVGG backbone network, and outputting the image features extracted by the RepVGG backbone network; The feature map of the detection head and the image features extracted by the RepVGG backbone network are input into the backbone network based on ShuffleNetv2, and the backbone network based on ShuffleNetv2 outputs the train track fastener missing recognition result.

2. The method according to claim 1, characterized in that The head detection and decoupling operation is performed on the train track fastener image data to be identified to obtain the features of the detected head, including: The image data of train track fasteners to be identified is collected by drones, and a 1×1 convolution is used to reduce the dimension of the train track fastener image data. Then, two 3×3 convolutions are used in the classification and regression branches respectively to perform head detection and decoupling operations on the train track fastener image data. Parameters are added on the basis of the original detection head to generate a feature map of the decoupled detection head.

3. The method according to claim 1 or 2, characterized in that: The step of inputting the feature map of the detected head into the RepVGG backbone network and outputting the image features extracted by the RepVGG backbone network comprises: The feature map of the detected head is input into the RepVGG backbone network. In the RepVGG backbone network, the 1×1 convolution kernel is converted into a 3×3 convolution kernel by padding. For the identity branch, the current convolution kernel parameter is set to 1, and the other convolution kernel parameters are set to 0. The identity branch is first converted into a 1×1 convolution kernel, and then converted into a 3×3 convolution kernel by the above padding operation. Finally, the 3×3 convolutions of the three output scales are added together by using the additivity of convolution to obtain the final convolution kernel. The RepVGG backbone network extracts features from the feature map of the detection head, and adjusts the feature map of the detection head to the size required by the network through standardization. During the training phase, each convolution layer of RepVGG consists of multiple parallel convolution paths. The 3x3 convolution branch is used to extract local spatial features, the 1x1 convolution branch is used to mix feature channels, and the identity mapping branch is used to maintain the residual connection of the features. The outputs of all branches are point-by-point additively fused, and nonlinearity is introduced through the ReLU activation function. The last layer of the RepVGG network outputs compressed image features, which capture high-level semantic information in the input image.

4. The method according to claim 2 or 3, characterized in that: The step of inputting the feature map of the detection head and the image features extracted by the RepVGG backbone network into a backbone network based on ShuffleNetv2, wherein the backbone network based on ShuffleNetv2 outputs a train track fastener missing recognition result, comprises: A backbone network based on ShuffleNetV2 is constructed according to four convolutional network design criteria, which include: (G1) when the number of input feature channels of the convolutional layer is equal to the number of output feature channels, the memory access time MAC is minimized; (G2) using grouped convolution with a large number of groups will increase MAC; (G3) fragmented parallel operations will reduce the model inference speed; (G4) considering the memory usage and time consumption caused by element-by-element operations; The feature map of the detected head and the image features extracted by the RepVGG backbone network are input into the backbone network based on ShuffleNetv2. The ShuffleNetv2 network linearly combines the features of different channels through point-by-point convolution, and splices the features from different networks together. The channels of the input feature map are divided into multiple groups, and each group performs a convolution operation separately. ShuffleNetv2 mixes the channels after convolution by group, so that the information between different groups interacts with each other. Each channel performs convolution separately to extract spatial features, and uses 1x1 convolution to combine the features of different channels to achieve channel information fusion. The convolution output is normalized to generate an optimized feature map; Perform subsequent missing fastener detection according to the optimized feature graph to obtain a missing train track fastener recognition result, which includes the detected fastener position, the missing fastener status, precision and recall information; The optimized feature map is input into the detection model, and the detection model decodes the optimized feature map to generate a detection frame and a category probability, and predicts the detection result of the fastener in each detection frame, calculates the confidence of each detection frame according to the category probability of each detection frame, filters out the detection frames with a confidence lower than a set threshold, and uses a non-maximum suppression algorithm to remove redundant detection frames to obtain a train track fastener missing recognition result, which includes the detected fastener position, the status of the fastener missing, the precision rate and the recall rate information.

Images