Track surface area identification method, device, equipment, and storage medium

Abstract

The present application provides a method, device, equipment, and storage medium for identifying a track surface area. The method includes: obtaining a track surface image; identifying a track surface area in the track surface image by using a pre-trained track surface area recognition model; wherein the track surface area recognition model is composed of a recognition module and an alignment module; the recognition module is constructed based on a YOLO model; the convolution layer of the recognition module is implemented using a depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the track surface image; and an alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module. The recognition module can reduce the amount of characteristic parameters and computational complexity of the track surface recognition process, thereby ensuring the real-time performance of the track surface recognition; the alignment module not only significantly improves the accuracy of the track surface area recognition, but also enables the track surface area recognition model to accurately recognize the track surface area under different lines.

Description

Technical Field

[0001] The present application relates to the field of rail transit technology, and in particular to a method, device, equipment, and storage medium for identifying a track surface area. Background Art

[0002] Rail adhesion refers to the roughness of the wheel-rail contact surface and whether there is a third medium between the wheel and rail, such as water, snow, fallen leaves and other substances. Under different working conditions, the available wheel-rail adhesion is different. If the same travel speed is used under different working conditions, it will inevitably cause the wheels of the rail vehicle to slide, which may cause wheel-rail scratches and loss of braking force in severe cases, greatly affecting driving safety.

[0003] The recognition accuracy of the track surface area is crucial for analyzing the track surface adhesion condition. If the recognition accuracy of the track surface area is poor, the algorithm may mistakenly identify the non-track surface area as the track surface area, or misjudge the real track surface area as the non-track surface area. This recognition error will directly affect the judgment of the track surface adhesion condition, which may lead to inaccurate track surface adhesion condition assessment and potential safety hazards. Improving the recognition accuracy of the track surface area can not only effectively improve the analysis of the track surface adhesion condition, but also provide more reliable safety protection for rail transit.

[0004] At present, there are two main situations for the recognition of track surface areas: one is that the recognition accuracy is high, but the speed cannot meet the real-time requirements; the other is that the recognition speed can reach real-time, but the accuracy is poor. However, when applied in the field of rail transportation, the recognition of track surface areas needs to maintain high accuracy and must meet the real-time requirements. Summary of the invention

[0005] In order to solve one of the above-mentioned technical defects, the present application provides a track surface area identification method, device, equipment, and storage medium.

[0006] In a first aspect, the present application provides a method for identifying a track surface area, the method comprising:

[0007] Acquire rail surface images;

[0008] Identify the rail surface area in the rail surface image by using a pre-trained rail surface area recognition model;

[0009] Among them, the track surface area recognition model consists of a recognition module and an alignment module;

[0010] The recognition module is built based on the YOLO model; the convolution layer of the recognition module is implemented using depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined based on the track surface image;

[0011] The alignment module is used to explicitly supervise the quality of the track surface mask obtained by the recognition module.

[0012] Optionally, the depthwise separable convolution includes a depthwise convolution operation and a pointwise convolution operation;

[0013] The deep convolution operation includes: performing convolution operations on each channel of the feature map of the rail surface image;

[0014] The point-by-point convolution operation includes fusing the convolution results of each channel through a 1×1 convolution kernel.

[0015] Optionally, when performing convolution operations on each channel separately, the convolution operations are not shared between different channels.

[0016] Optionally, the convolution results of each channel are fused through a 1×1 convolution kernel, including:

[0017] The convolution results of each channel are weighted summed at each pixel position through a 1×1 convolution kernel.

[0018] Optionally, the number of prototypes in the convolutional layer of the recognition module is generated based on the global features through a fully connected layer using a ReLU activation function;

[0019] Among them, the global feature is the feature map of the track surface image extracted through the global average pooling layer.

[0020] Optionally, the alignment module consists of three cascaded depthwise separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

[0021] Optionally, the quality of the track surface area mask is the product of the score of the track surface area mask given by the recognition module and the category confidence.

[0022] In a second aspect of the present application, a rail surface area identification device is provided, the device comprising:

[0023] An acquisition module, used for acquiring rail surface images;

[0024] A recognition module, used for recognizing the rail surface area in the rail surface image acquired by the acquisition module through a pre-trained rail surface area recognition model;

[0025] Among them, the track surface area recognition model consists of a recognition module and an alignment module;

[0026] The recognition module is built based on the YOLO model; the convolution layer of the recognition module is implemented using depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined based on the track surface image;

[0027] The alignment module is used to explicitly supervise the quality of the track surface mask obtained by the recognition module.

[0028] In a third aspect of the present application, an electronic device is provided, including:

[0029] Memory;

[0030] Processor; and

[0031] Computer programs;

[0032] The computer program is stored in the memory and is configured to be executed by the processor to implement the method as described in the first aspect above.

[0033] In a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement the method described in the first aspect above.

[0034] The present application provides a method, device, equipment, and storage medium for identifying a track surface area. The method includes: obtaining a track surface image; identifying a track surface area in the track surface image by using a pre-trained track surface area recognition model; wherein the track surface area recognition model is composed of a recognition module and an alignment module; the recognition module is constructed based on a YOLO model; the convolution layer of the recognition module is implemented using a depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the track surface image; and an alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module. The recognition module can reduce the amount of characteristic parameters and computational complexity of the track surface recognition process, thereby ensuring the real-time performance of the track surface recognition; the alignment module not only significantly improves the accuracy of the track surface area recognition, but also enables the track surface area recognition model to accurately recognize the track surface area under different lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0036] Figure 1 A schematic diagram of a flow chart of a rail surface area identification method provided in an embodiment of the present application;

[0037] Figure 2 A schematic diagram of the structure of a track surface area identification model provided in an embodiment of the present application;

[0038] Figure 3 A schematic diagram of the structure of a convolutional layer of a recognition module provided in an embodiment of the present application;

[0039] Figure 4 A schematic diagram of the structure of an alignment module provided in an embodiment of the present application;

[0040] Figure 5 A schematic diagram of the structure of a rail surface area identification device provided in an embodiment of the present application;

[0041] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the technical solutions and advantages in the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0043] In the process of realizing this application, the inventor found that there are two main situations for the current recognition of track surface areas: one is that the recognition accuracy is high, but the speed cannot meet the real-time requirements; the other is that the recognition speed can reach real-time, but the accuracy is poor. However, when applied in the field of rail transportation, the recognition of track surface areas needs to maintain high accuracy and must meet the real-time requirements.

[0044] In response to the above problems, the present application provides a method, device, equipment, and storage medium for identifying a track surface area. The method includes: obtaining a track surface image; identifying a track surface area in the track surface image through a pre-trained track surface area recognition model; wherein the track surface area recognition model is composed of a recognition module and an alignment module; the recognition module is constructed based on the YOLO model; the convolution layer of the recognition module is implemented using a depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the track surface image; and an alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module. The recognition module can reduce the amount of characteristic parameters and computational complexity of the track surface recognition process, thereby ensuring the real-time nature of the track surface recognition; the alignment module not only significantly improves the accuracy of the track surface area recognition, but also enables the track surface area recognition model to accurately recognize the track surface area under different lines.

[0045] See also Figure 1 This embodiment provides a method for identifying a track surface area. The implementation process of the method is as follows:

[0046] 101, obtaining a rail surface image.

[0047] 102 , identifying a rail surface area in the rail surface image by using a pre-trained rail surface area recognition model.

[0048] Among them, the track surface area recognition model consists of a recognition module and an alignment module, such as Figure 2 shown. Figure 2 The AlignNet in is the alignment module.

[0049] 1. Identification module

[0050] The recognition module is built based on the YOLO model.

[0051] Among them, YOLO is a network model that has been widely used in the industry in recent years. It was originally designed for target detection tasks, and the output module can be flexibly replaced according to task requirements to adapt to different application scenarios. YOLO's Backbone network focuses on the extraction of multi-scale feature maps, while the Neck network is used to effectively fuse these feature maps to enhance the detection ability of the model. Due to its relatively simple network structure, YOLO performs well in reasoning efficiency and detection accuracy, and is especially suitable for scenarios with high real-time requirements. In addition, YOLO's modular design facilitates further optimization and adjustment, so that it can also achieve good performance on different devices (such as mobile devices, edge devices), and is widely used in industrial detection, autonomous driving, security monitoring and other fields.

[0052] The YOLACT algorithm implements instance segmentation through two parallel sub-network branches. Among them, the Prediction Head branch generates the category confidence, position regression parameters and mask coefficients of each anchor box, while the ProtoNet branch generates a set of prototype masks. By multiplying the prototype mask with the target mask coefficient, the final mask of each target in the image can be obtained. YOLACT uses a single-stage network framework. Although it reduces the amount of calculation to a certain extent, it still appears to be computationally intensive on the vehicle-mounted edge device and is difficult to meet real-time requirements. In addition, the problem of insufficient accuracy is prone to occur when identifying the track surface of different lines, which shows that the feature extraction effect of the YOLACT algorithm has not yet reached a sufficient level, and it is difficult to achieve accurate recognition on existing data sets.

[0053] When constructing the recognition module, this embodiment integrates YOLO and YOLACT, and improves the integrated model to meet the requirements of track surface area recognition in the field of rail transit, so that the track surface area recognition model can accurately identify the track surface area on different lines, which not only improves the reliability of recognition, but also provides important guarantee for the safe operation of trains.

[0054] The specific fusion process is: based on the YOLO model, the object detection head of the YOLO model is replaced with the instance segmentation head of the YOLACT model. The parts that are improved after fusion are as follows:

[0055] 1) The convolution layer of the recognition module is implemented using Depth Separable Convolution (DSC).

[0056] In specific implementation, the depthwise separable convolution consists of two steps, which can significantly reduce the amount of calculation and parameters while maintaining the feature expression ability. The first step is the depthwise convolution operation, and the second step is the pointwise convolution operation. Therefore, the depthwise separable convolution (DSC) includes the depthwise convolution operation and the pointwise convolution operation.

[0057] Among them, the deep convolution operation includes: performing convolution operations on each channel of the feature map of the rail surface image. In addition, when the convolution operation is performed on each channel, the convolution operation is not shared between different channels. During the deep convolution operation, the convolution operation is performed on each channel of the feature map of the rail surface image, each channel uses an independent convolution kernel, and the convolution operation is not shared between different channels, so as to obtain a separate convolution result for each channel. This step focuses on feature extraction in the spatial dimension.

[0058] The point-by-point convolution operation includes: fusing the convolution results of each channel through a 1×1 convolution kernel. For example, through a 1×1 convolution kernel, the convolution results of each channel are weighted summed at each pixel position. During the point-by-point convolution operation, the convolution results of each channel are fused using a 1×1 convolution kernel. This step does not translate the convolution kernel in the spatial dimension, but weighted sums the features of each channel at each pixel position to achieve information integration between channels. This convolution method not only retains the expressive power of traditional convolution, but also greatly reduces the computational complexity and the number of parameters, generally achieving a 9 to 10 times reduction effect, depending on the size of the convolution kernel and the number of input and output channels.

[0059] The convolutional layer of the recognition module uses depthwise separable convolution (DSC), which performs well in compressing and accelerating the track area recognition model. It is particularly suitable for application in scenarios with limited computing resources such as mobile devices or edge computing.

[0060] The convolution layer of the ProtoNet module in the YOLACT model adopts a standard convolution layer. After the target detection head of the YOLO model is replaced with the YOLACT model in the recognition module of this embodiment, the convolution layer of the ProtoNet module is improved to a deep separable convolution (DSC), which can reduce the computational cost while maintaining the recognition accuracy.

[0061] This improvement actually replaces all the standard convolutional layers in the ProtoNet module in the replaced YOLACT model with depth-to-depth separable convolutions, which reduces the computational cost while maintaining recognition accuracy.

[0062] 2) The number of prototypes in the convolutional layer of the recognition module is determined based on the track surface image.

[0063] For example, the number of prototypes in the convolutional layer of the recognition module is generated based on the global features through the fully connected layer using the ReLU activation function.

[0064] Among them, the global feature is the feature map of the track surface image extracted through the global average pooling layer.

[0065] The number of prototypes of the convolution layer in the YOLACT model is fixed. After the target detection head of the YOLO model is replaced with the YOLACT model in the recognition module of this embodiment, the number of prototypes of the convolution layer is improved to be dynamic, which can improve the adaptability of the track surface area recognition model in different scenarios.

[0066] like Figure 2 The DDSC-ProtoNet in the figure is the improved ProtoNet, and its structure is as follows Figure 3 As shown. Compared with the standard convolution layer in ProtoNet, the convolution layer of DDSC-ProtoNet is implemented through depthwise separable convolution (DSC), which first generates the convolution result of a single channel of the feature map through a depthwise convolution operation, and then aggregates the convolution results of multiple channels through a pointwise convolution operation to obtain the final convolution layer output. The depthwise separable convolution adopted by DDSC-ProtoNet is a lightweight convolution operation that can significantly reduce the amount of feature parameters and computational complexity, as well as the risk of overfitting, so that DDSC-ProtoNet can improve processing efficiency while maintaining high accuracy, making it more suitable for real-time applications.

[0067] Unlike the fixed number of prototypes in ProtoNet, the number of prototypes in DDSC-ProtoNet is dynamically generated. Specifically, the number of prototypes in DDSC-ProtoNet is automatically determined by the dynamic branch according to the content of the feature map of the track surface image. That is, the feature map of the track surface image first extracts global features through the global average pooling layer, then is processed by the fully connected layer, and the ReLU activation function is used to generate the number of prototypes that adapt to the current track surface image requirements. Compared with the fixed number of prototypes, the dynamic number of prototypes in this embodiment has the following three advantages: (1) Strong scene adaptability. The number of dynamic prototypes can be adjusted according to the complexity of the feature map, so that it can better adapt to the scene. For example, in a simple scene, only a few prototypes may be needed to represent the main target, while in a complex scene, more prototypes are needed to capture details and diverse target shapes; (2) The track surface area recognition model has a high recognition accuracy. The dynamic adjustment of the number of prototypes enables the track surface area recognition model to generate more masks in complex images, thereby improving the recognition accuracy of the track surface area recognition model; (3) High computational efficiency. By dynamically adjusting the number of prototypes according to the input, the amount of computation can be reduced in simple scenes, thereby improving the inference speed and efficiency. In situations where computational resources are limited, dynamic prototyping can help optimize the computational requirements of a model.

[0068] After the recognition module integrates YOLO and YOLACT, the convolution layer is improved to be implemented through deep separable convolution (DSC) to address the poor real-time performance of YOLACT. This can effectively reduce the number of parameters of the track surface area recognition model while maintaining the same accuracy, thereby significantly improving the recognition speed of the track surface area recognition model. In addition, in response to the problem of insufficient scene adaptability of YOLACT, the number of prototypes in the convolution layer is improved to a dynamic number, which can further enhance the scene adaptability of the track surface area recognition model.

[0069] 2. Alignment module

[0070] Since there is a misalignment between the category confidence of the detection box generated by the Prediction Head branch and the quality of the generated mask, that is, a high category confidence does not necessarily mean that the generated mask quality is also high, the recognition module cannot explicitly supervise the quality of the target mask (i.e., the track surface area mask) after fusing and improving YOLO and YOLACT. This inconsistency may lead to poor detection results for some categories in practical applications, affecting the overall performance of the model. Therefore, an alignment module (such as Figure 2 AlignNet in the recognition module can establish a correlation between the confidence generated by the PredictionHead branch in the recognition module and the quality of the target mask, explicitly supervising the quality of the generated target mask.

[0071] Therefore, the alignment module is used to explicitly supervise the quality of the track surface region mask obtained by the recognition module.

[0072] For example, the alignment module consists of three cascaded depthwise separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

[0073] The quality of the track surface area mask is the product of the score of the recognition module for the track surface area mask and the category confidence.

[0074] like Figure 2 The AlignNet structure in Figure 4 As shown in the figure, the AlignNet module consists of three cascaded depth-separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer. AlignNet multiplies the score of the track surface area mask output by the recognition module with the category confidence as the score of each mask. This scoring mechanism effectively establishes a close connection between the category confidence and the generated mask, ensuring that high-confidence categories correspond to high-quality masks.

[0075] The alignment module can solve the problem of lack of supervision of the track area mask generated after the fusion of YOLO and YOLACT. The alignment module can explicitly supervise the generation quality of the alignment module.

[0076] In addition, the alignment module maintains computational efficiency through the lightweight characteristics of deep separable convolution. In this way, the alignment module not only significantly improves the accuracy of mask generation, but also enhances the robustness and adaptability of the track surface area recognition model in instance segmentation tasks, enabling it to maintain good performance in complex scenarios. This clear supervision mechanism provides a more solid foundation for the subsequent training and tuning of the track surface area recognition model, and helps to further promote the development of instance segmentation technology.

[0077] The present embodiment provides a method for identifying a track surface area, which includes obtaining a track surface image; identifying a track surface area in the track surface image through a pre-trained track surface area identification model; wherein the track surface area identification model is composed of a recognition module and an alignment module; the recognition module is constructed based on a YOLO model; the convolution layer of the recognition module is implemented using a depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the track surface image; and the alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module. The recognition module can reduce the amount of characteristic parameters and computational complexity of the track surface identification process, thereby ensuring the real-time performance of the track surface identification; the alignment module not only significantly improves the accuracy of the track surface area identification, but also enables the track surface area identification model to accurately identify the track surface area under different lines.

[0078] Based on the same inventive concept of the rail surface area recognition method, this embodiment provides a rail surface area recognition device, see Figure 5 , the device comprises:

[0079] The acquisition module 501 is used to acquire the rail surface image.

[0080] The recognition module 502 is used to recognize the rail surface area in the rail surface image acquired by the acquisition module 501 by using a pre-trained rail surface area recognition model.

[0081] Among them, the track surface area recognition model consists of a recognition module and an alignment module.

[0082] The recognition module is built based on the YOLO model. The convolution layer of the recognition module is implemented using depthwise separable convolution. The number of prototypes in the convolution layer of the recognition module is determined based on the track surface image.

[0083] The alignment module is used to explicitly supervise the quality of the track surface mask obtained by the recognition module.

[0084] Among them, the depth-wise separable convolution includes depth-wise convolution operation and point-by-point convolution operation.

[0085] The deep convolution operation includes: performing convolution operations on each channel of the feature map of the track surface image respectively.

[0086] The point-by-point convolution operation includes fusing the convolution results of each channel through a 1×1 convolution kernel.

[0087] Among them, when the convolution operation is performed on each channel separately, the convolution operation is not shared between different channels.

[0088] Among them, the convolution results of each channel are fused through a 1×1 convolution kernel, including:

[0089] The convolution results of each channel are weighted summed at each pixel position through a 1×1 convolution kernel.

[0090] Among them, the number of prototypes in the convolutional layer of the recognition module is generated based on the global features through the fully connected layer using the ReLU activation function.

[0091] Among them, the global feature is the feature map of the track surface image extracted through the global average pooling layer.

[0092] Among them, the alignment module consists of three cascaded depth-separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

[0093] The quality of the track surface area mask is the product of the score of the recognition module for the track surface area mask and the category confidence.

[0094] The device provided in this embodiment identifies the rail surface area in the rail surface image through a pre-trained rail surface area recognition model, wherein the recognition module can reduce the number of characteristic parameters and the computational complexity of the rail surface recognition process, thereby ensuring the real-time performance of the rail surface recognition; the alignment module not only significantly improves the accuracy of the rail surface area recognition, but also enables the rail surface area recognition model to accurately recognize the rail surface area under different lines.

[0095] Based on the same inventive concept of the rail surface area recognition method, this embodiment provides an electronic device, the electronic device is as follows Figure 6 As shown, it includes: a memory 601, a processor 602, and a computer program.

[0096] The computer program is stored in the memory 601 and is configured to be executed by the processor 602 to implement the above-mentioned rail surface area identification method.

[0097] Specifically,

[0098] Get rail surface images.

[0099] The track surface area in the track surface image is identified by using a pre-trained track surface area recognition model.

[0100] Among them, the track surface area recognition model consists of a recognition module and an alignment module.

[0101] The recognition module is built based on the YOLO model. The convolution layer of the recognition module is implemented using depthwise separable convolution. The number of prototypes in the convolution layer of the recognition module is determined based on the track surface image.

[0102] The alignment module is used to explicitly supervise the quality of the track surface mask obtained by the recognition module.

[0103] Optionally, the depthwise separable convolution includes a depthwise convolution operation and a pointwise convolution operation.

[0104] The deep convolution operation includes: performing convolution operations on each channel of the feature map of the track surface image respectively.

[0105] The point-by-point convolution operation includes fusing the convolution results of each channel through a 1×1 convolution kernel.

[0106] Optionally, when performing convolution operations on each channel separately, the convolution operations are not shared between different channels.

[0107] Optionally, the convolution results of each channel are fused through a 1×1 convolution kernel, including:

[0108] The convolution results of each channel are weighted summed at each pixel position through a 1×1 convolution kernel.

[0109] Optionally, the number of prototypes in the convolutional layer of the recognition module is generated based on the global features through a fully connected layer using a ReLU activation function.

[0110] Among them, the global feature is the feature map of the track surface image extracted through the global average pooling layer.

[0111] Optionally, the alignment module consists of three cascaded depthwise separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

[0112] Optionally, the quality of the track surface area mask is the product of the score of the track surface area mask given by the recognition module and the category confidence.

[0113] The electronic device provided by the present embodiment has a computer program executed by a processor to identify the rail surface area in the rail surface image through a pre-trained rail surface area recognition model, wherein the recognition module can reduce the amount of characteristic parameters and the computational complexity of the rail surface recognition process, thereby ensuring the real-time performance of the rail surface recognition; the alignment module not only significantly improves the accuracy of the rail surface area recognition, but also enables the rail surface area recognition model to accurately recognize the rail surface area under different lines.

[0114] Based on the same inventive concept of the rail surface region identification method, this embodiment provides a computer-readable storage medium, and a computer program is stored thereon. The computer program is executed by a processor to implement the rail surface region identification method.

[0115] Specifically,

[0116] Get rail surface images.

[0117] The track surface area in the track surface image is identified by using a pre-trained track surface area recognition model.

[0118] Among them, the track surface area recognition model consists of a recognition module and an alignment module.

[0119] The recognition module is built based on the YOLO model. The convolution layer of the recognition module is implemented using depthwise separable convolution. The number of prototypes in the convolution layer of the recognition module is determined based on the track surface image.

[0120] The alignment module is used to explicitly supervise the quality of the track surface mask obtained by the recognition module.

[0121] Optionally, the depthwise separable convolution includes a depthwise convolution operation and a pointwise convolution operation.

[0122] The deep convolution operation includes: performing convolution operations on each channel of the feature map of the track surface image respectively.

[0123] The point-by-point convolution operation includes fusing the convolution results of each channel through a 1×1 convolution kernel.

[0124] Optionally, when performing convolution operations on each channel separately, the convolution operations are not shared between different channels.

[0125] Optionally, the convolution results of each channel are fused through a 1×1 convolution kernel, including:

[0126] The convolution results of each channel are weighted summed at each pixel position through a 1×1 convolution kernel.

[0127] Optionally, the number of prototypes in the convolutional layer of the recognition module is generated based on the global features through a fully connected layer using a ReLU activation function.

[0128] Among them, the global feature is the feature map of the track surface image extracted through the global average pooling layer.

[0129] Optionally, the alignment module consists of three cascaded depthwise separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

[0130] Optionally, the quality of the track surface area mask is the product of the score of the track surface area mask given by the recognition module and the category confidence.

[0131] The computer readable storage medium provided in this embodiment has a computer program thereon executed by a processor to identify the rail surface area in the rail surface image through a pre-trained rail surface area recognition model, wherein the recognition module can reduce the amount of characteristic parameters and the computational complexity of the rail surface recognition process, thereby ensuring the real-time performance of the rail surface recognition; the alignment module not only significantly improves the accuracy of the rail surface area recognition, but also enables the rail surface area recognition model to accurately recognize the rail surface area under different lines.

[0132] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0133] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0134] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0136] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0137] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying a rail surface area, characterized in that: The method comprises: Acquire rail surface images; identifying the rail surface area in the rail surface image by using a pre-trained rail surface area recognition model; Among them, the track surface area recognition model consists of a recognition module and an alignment module; The recognition module is constructed based on the YOLO model; the convolution layer of the recognition module is implemented by depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the rail surface image; The alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module.

2. The method according to claim 1, characterized in that The depth-wise separable convolution includes a depth-wise convolution operation and a point-wise convolution operation; The deep convolution operation includes: performing convolution operations on each channel of the feature map of the track surface image respectively; The point-by-point convolution operation includes fusing the convolution results of each channel through a 1×1 convolution kernel.

3. The method according to claim 2, characterized in that When convolution operations are performed on each channel separately, convolution operations are not shared between different channels.

4. The method according to claim 2, characterized in that: The convolution results of each channel are fused by the 1×1 convolution kernel, including: The convolution results of each channel are weighted summed at each pixel position through a 1×1 convolution kernel.

5. The method according to claim 1, characterized in that The number of prototypes in the convolutional layer of the recognition module is generated based on global features through a fully connected layer using a ReLU activation function; The global feature is a feature map of the track surface image extracted through a global average pooling layer.

6. The method according to claim 1, characterized in that The alignment module consists of three cascaded depth-wise separable convolutional layers, a batch normalization layer, a ReLU activation function, and a downsampling layer.

7. The method according to claim 6, characterized in that The quality of the rail surface area mask is the product of the score of the rail surface area mask given by the recognition module and the category confidence.

8. A rail surface area identification device, characterized in that: The device comprises: An acquisition module, used for acquiring rail surface images; A recognition module, used for recognizing the rail surface area in the rail surface image acquired by the acquisition module through a pre-trained rail surface area recognition model; Among them, the track surface area recognition model consists of a recognition module and an alignment module; The recognition module is constructed based on the YOLO model; the convolution layer of the recognition module is implemented by depthwise separable convolution; the number of prototypes in the convolution layer of the recognition module is determined according to the rail surface image; The alignment module is used to explicitly supervise the quality of the track surface area mask obtained by the recognition module.

9. An electronic device, characterized in that: include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon; the computer program is executed by a processor to implement the method according to any one of claims 1 to 7.

Images