Rail transit small target recognition method based on two-stage residual diffusion model

Abstract

The application discloses a rail transit small target recognition method based on a two-stage residual diffusion model, and comprises the following steps: acquiring multi-modal data to be recognized; inputting the multi-modal data to be recognized into a trained first residual diffusion model for processing, predicting key area features, and reconstructing the multi-modal data to be recognized; inputting the reconstructed multi-modal data to be recognized and the multi-modal data to be recognized into a trained second residual diffusion model for processing, and obtaining a small target recognition result; wherein the first residual diffusion model and the second residual diffusion model are trained in a joint training mode, pre-constructed samples are used as training data, pre-generated pseudo-true values are used as sample labels, and the initial first residual diffusion model and the second residual diffusion model are trained to obtain the trained first residual diffusion model and the trained second residual diffusion model. The application can effectively overcome the defects that small target features are weak and easy to be submerged.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, intelligent transportation and rail transit technology, and specifically relates to a method for small target recognition in rail transit based on a two-stage residual diffusion model. Background Technology

[0002] The safe operation of rail transit relies on accurate perception of the track area and its surrounding environment. Among these, the identification of small targets is a key link in ensuring safe operation, such as pedestrians, small obstacles, and foreign objects on the track. However, the rail transit environment is complex and changeable, with various factors such as changes in lighting, severe weather, sensor noise, and motion blur, which seriously affect the performance of traditional visual perception systems.

[0003] The main problems currently facing small target recognition in rail transit include: First, the target is small in scale and low in resolution; in long-distance monitoring or wide-angle field of view, the pixel ratio of small targets is extremely small, making feature extraction difficult; Second, severe environmental degradation; weather conditions such as rain, snow, fog and haze cause image quality to decline, blurring target boundaries and reducing contrast; Third, multimodal data fusion is difficult; data from multiple sources such as visible light, infrared, radar, and lidar face challenges in spatiotemporal alignment and feature complementarity.

[0004] Most existing methods rely on single-modal deep convolutional networks for target detection, which leads to a sharp decline in performance under complex environments. A few multimodal methods employ simple feature concatenation or decision-level fusion, which fails to fully utilize the complementary information between multiple modalities and lacks robustness to sensor degradation. Therefore, there is an urgent need to propose a method for small target recognition in rail transit systems to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a method for identifying small targets in rail transit based on a two-stage residual diffusion model. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for identifying small targets in rail transit based on a two-stage residual diffusion model, comprising: Acquire the multimodal data to be identified; the multimodal data to be identified includes small targets; The multimodal data to be identified is input into a trained first residual diffusion model for processing to predict key region features and reconstruct the multimodal data to be identified. The reconstructed multimodal data and the multimodal data to be identified are then input into a trained second residual diffusion model for processing to obtain the identification result of the small target. The first residual diffusion model and the second residual diffusion model are trained by joint training. The pre-constructed samples are used as training data and the pre-generated pseudo-true values ​​are used as the labels of the samples. The initial first residual diffusion model and the second residual diffusion model are trained to obtain the trained first residual diffusion model and the trained second residual diffusion model.

[0006] The beneficial effects of this invention are: This invention provides a method for identifying small targets in rail transit based on a two-stage residual diffusion model. The method employs a two-stage residual diffusion framework of reconstruction and identification, which can not only restore the details of the image, but more importantly, actively predict the features of key areas to form a structural guide for subsequent identification. In the second stage, the reconstructed data is spliced ​​with the original data and refined through the second residual diffusion model to obtain the identification result, effectively overcoming the defects of weak features and easy obscuration of small targets.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0008] Figure 1 This is a flowchart of a method for identifying small targets in rail transit based on a two-stage residual diffusion model, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a method for identifying small targets in rail transit based on a two-stage residual diffusion model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of a trained first residual diffusion model or a trained second residual diffusion model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a spatial information recovery module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a fusion module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a residual prediction network provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0010] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart of a method for identifying small targets in rail transit based on a two-stage residual diffusion model, provided in an embodiment of the present invention. Figure 2This is a schematic diagram of a small target recognition method for rail transit based on a two-stage residual diffusion model provided by an embodiment of the present invention. The method for small target recognition in rail transit based on a two-stage residual diffusion model provided by the present invention includes: S101. Obtain the multimodal data to be identified; wherein, the multimodal data to be identified includes small targets.

[0011] Specifically, in this embodiment, the multimodal data to be identified may include visible light images. Infrared images LiDAR point cloud and millimeter wave data The multimodal data is synchronized in time and registered in space to form aligned data pairs; the small targets in the multimodal data to be identified can be pedestrians, small obstacles or foreign objects on the track.

[0012] S102. The multimodal data to be identified is input into the trained first residual diffusion model for processing to predict key region features and reconstruct the multimodal data to be identified; the reconstructed multimodal data to be identified and the multimodal data to be identified are input into the trained second residual diffusion model for processing to obtain the identification result of the small target; wherein, The first residual diffusion model and the second residual diffusion model are trained by joint training. The pre-constructed samples are used as training data and the pre-generated pseudo-true values ​​are used as the labels of the samples. The initial first residual diffusion model and the second residual diffusion model are trained to obtain the trained first residual diffusion model and the trained second residual diffusion model.

[0013] Specifically, in this embodiment, please refer to Figure 3 , Figure 3 This is a schematic diagram of the architecture of a trained first residual diffusion model or a trained second residual diffusion model provided in an embodiment of the present invention. The architecture of the trained first residual diffusion model and the trained second residual diffusion model is the same, both including: a backbone network, an augmented encoder, a residual diffusion model, and a decoder; wherein, The backbone network in the first residual diffusion model is used to extract multi-scale features from the multimodal data to be identified, thus obtaining the first multi-scale features. The enhancement encoder in the first residual diffusion model is used to perform multi-scale compression and global feature enhancement on the first multi-scale feature, and then fuses the enhanced global feature with the first multi-scale feature to obtain the first enhanced feature; The residual diffusion model in the first residual diffusion model is used to predict the first enhanced feature to obtain the first structural feature; The decoder in the first residual diffusion model is used to reconstruct the first structural features to obtain the reconstructed multimodal data to be identified. The backbone network in the second residual diffusion model is used to extract multi-scale features from the spliced ​​data of the multimodal data to be identified and the reconstructed multimodal data to be identified, thus obtaining the second multi-scale features. The enhanced encoder in the second residual diffusion model is used to perform multi-scale compression and global feature enhancement on the second multi-scale features, and then fuses the enhanced global features with the second multi-scale features to obtain the second enhanced features. The residual diffusion model in the second residual diffusion model is used to predict the second enhanced feature to obtain the second structural feature; In the second residual diffusion model, the decoder is used to reconstruct the second structural features to obtain the recognition result of small targets.

[0014] It should be noted that in this embodiment, the first residual diffusion model is a structural feature diffusion model, and the second residual diffusion model is a small target recognition diffusion model. Both are constructed based on a backbone network, an enhanced encoder, a residual diffusion model, and a decoder. In the first residual diffusion model, the structural feature map of the small target is learned and predicted. This method highlights regions in an image that may contain small objects, such as high-gradient areas, areas of abrupt motion changes, and areas with complex textures. The input is multimodal data. , Represents a point cloud projection. To represent the radar density map, a backbone network is used to extract multi-scale features, an augmented encoder is used to obtain global augmented features, a residual diffusion model is used to predict the global augmented features, and then a decoder is used to reconstruct the reconstructed multimodal data to be identified; in the second residual diffusion model, the structural feature map generated in the first stage is used. As a conditional guide, it enables accurate detection and classification of small targets, with multimodal data as input. and structural feature diagram The splicing features are obtained by using a backbone network to extract multi-scale features, an augmented encoder to obtain global augmented features, a residual diffusion model to predict the global augmented features, and then a decoder to reconstruct the small target to obtain the recognition result, i.e., heat map or category map. It should be noted that the heat map is finally thresholded and non-maximum suppression is applied, and combined with the category map, the target bounding box, category and confidence score are output.

[0015] Furthermore, in this embodiment, the enhanced encoder includes a multi-scale enhancement module and a spatial information recovery module; wherein, The enhancement module is used to fuse and enhance multi-scale features to obtain enhanced global features; The spatial information recovery module is used to fuse enhanced global features and multi-scale features to obtain enhanced features.

[0016] Furthermore, in this embodiment, please refer to Figure 4 , Figure 4 This is a schematic diagram of a spatial information recovery module provided in an embodiment of the present invention. The spatial information recovery module includes a first upsampling module, a second upsampling module, a third upsampling module, a first downsampling module, a convolution module, a first fusion module, and a second fusion module; wherein, The enhanced global features are downsampled using the first downsampling module to obtain the first feature; The convolution module is used to perform a convolution operation on the feature concatenated from the first feature and the third scale feature to obtain the second feature; The first upsampling module upsamples the second feature to obtain the third feature; the third feature, the second scale feature, and the enhanced global feature are concatenated to obtain the fourth feature. The fourth feature is fused using the first fusion module to obtain the fifth feature; The fifth feature is upsampled using the second upsampling module to obtain the sixth feature; The enhanced global features are upsampled using the third upsampling module to obtain the seventh feature; The seventh feature, the first scale feature, and the sixth feature are concatenated to obtain the eighth feature; the eighth feature is fused using the second fusion module to obtain the ninth feature; the second feature, the fifth feature, and the ninth feature are concatenated to form an enhanced feature; wherein the second feature is used as a low-resolution feature, the fifth feature is used as a medium-resolution feature, and the ninth feature is used as a high-resolution feature.

[0017] Further, please see Figure 5 , Figure 5 This is a schematic diagram of a fusion module provided in an embodiment of the present invention. In this embodiment, the first fusion module and the second fusion module have the same structure, both including a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer; wherein, The first convolutional layer is used to perform a convolution operation on the input features to obtain the first convolutional features; The input features are convolved using a second convolutional layer to obtain the second convolutional features; The third convolutional layer is used to perform a convolution operation on the first convolutional features to obtain the third convolutional features; The fourth convolutional layer is used to perform a convolution operation on the third convolutional layer to obtain the fourth convolutional feature; The fifth convolutional layer is used to perform a convolution operation on the first convolutional features to obtain the fifth convolutional features; The fifth convolutional feature is added element-wise to the fourth convolutional feature to obtain the sixth convolutional feature; The sixth convolutional feature is added element-wise to the second convolutional feature to obtain the output feature.

[0018] Furthermore, in this embodiment, the processing of the multi-scale enhancement module includes: downsampling the first-scale features and the second-scale features through bilinear interpolation to obtain the first downsampled features and the second downsampled features; concatenating the first downsampled features and the second downsampled features in the channel dimension to obtain a feature map; adjusting the number of channels of the feature map using a 1×1 convolutional layer to obtain the adjusted features; processing the adjusted features using a channel attention mechanism; and then fusing them using a 3×3 convolutional layer to obtain the enhanced global features.

[0019] Specifically, in this embodiment, the residual diffusion model includes a conditional mapping network and a residual prediction network, wherein the residual prediction network adopts a U-Net structure; wherein, A conditional mapping network is used to map the enhanced features to preliminary estimated features; Using multi-scale resolution features as conditions, a residual prediction network is used to denoise the noise of random sampling and predict clean residuals. The clean residuals are then intersected with the preliminary estimated features to obtain the output of the residual diffusion model.

[0020] In this embodiment, the residual of the residual prediction network is defined as... , This represents a shallow feature extractor. The diffusion process takes place in the residual space, and the goal of training the residual diffusion model is to minimize the predicted structural feature map. Compared with the real structural feature map The differences between them are addressed by generating the true structural feature map through a combination of methods including edge detection, optical flow motion regions, and texture saliency. The forward diffusion process of the residual prediction network is represented as follows: ; in, Indicates conditional input. Representing the residual components, DDIM deterministic sampling is used to accelerate inference.

[0021] Further, please see Figure 6 , Figure 6This is a schematic diagram of a residual prediction network provided in an embodiment of the present invention. In this embodiment, the residual prediction network includes an upsampling module, a bottleneck module, and a downsampling module. The upsampling module includes a first residual module, a second residual module, and a third residual module, each of which includes two residual blocks. The bottleneck module includes a third fusion module. The downsampling module includes a fourth residual module, a fifth residual module, and a sixth residual module, each of which includes two residual modules. The fourth residual module further includes a cross-attention module, and the sixth residual module further includes an AdaIN module. The first residual module is used to process the randomly sampled noise and time step embedding vector to obtain the first residual feature; The second residual module is used to process the first residual feature to obtain the second residual feature; The third residual module is used to process the first residual feature to obtain the third residual feature; The bottleneck module is used to fuse the third residual feature with the low-resolution feature to obtain the fused feature; The fourth residual module is used to process the fused features to obtain the fourth residual features. A cross-attention module is used to fuse the medium-resolution feature with the fourth residual feature to obtain the second fused feature; the second fused feature is then added element-wise with the second residual feature to obtain the third fused feature. The fifth residual module is used to process the third fusion feature to obtain the fifth residual feature. The fifth residual feature is then added element-wise to the first residual feature to obtain the fourth fusion feature. The sixth residual module is used to process the fourth fusion feature to obtain the sixth residual feature; Based on the high-resolution features, the AdaIN module is used to fine-tune the sixth residual feature to obtain the output of the residual prediction network.

[0022] It should be noted that the bottleneck module is used to fuse the third residual feature with the low-resolution feature, that is, to concatenate the third residual feature with the low-resolution feature, and then adjust the number of channels through a convolutional layer to obtain the fused feature; the cross-attention module is used to fuse the medium-resolution feature and the fourth residual feature, that is, to use the fourth residual feature as the query and the medium-resolution feature as the key and value; the AdaIN module is used to fine-tune the sixth residual feature, that is, the high-resolution feature is used to generate a set of scaling and encoding parameters through a convolutional network, which is the same as the feature channel number of the corresponding layer of the U-Net decoder, and then the U-Net feature is modulated, and finally the output of the residual prediction network is output.

[0023] In summary, the small target identification method for rail transit based on a two-stage residual diffusion model provided by this invention has the following beneficial effects: First, this invention proposes a two-stage collaborative architecture to achieve high-precision recognition of small targets.

[0024] This invention employs a two-stage residual diffusion framework of reconstruction and recognition, which not only restores the details of the image, but more importantly, it can actively predict the features of key regions, forming a structural guide for subsequent recognition. In the second stage, the reconstructed data is stitched together with the original data, and the recognition is refined through the second residual diffusion model to obtain the recognition result, effectively overcoming the defects of weak features and easy obscuration of small targets.

[0025] Second, this invention proposes a residual diffusion mechanism that balances generation quality and inference efficiency.

[0026] This invention introduces a conditional mapping network into the diffusion model, which quickly maps enhanced features to preliminary estimates, allowing the diffusion process to focus on the learning target residuals rather than the complete distribution, thus greatly reducing the complexity of the generation task. The residual prediction network adopts the U-Net architecture and directly predicts clean residuals. Combined with the DDIM deterministic sampling strategy, the number of inference steps is compressed from thousands of steps in traditional diffusion models to 20-50 steps, with a single frame processing time of less than 100ms. While maintaining high-fidelity generation quality, this allows the diffusion model to meet the stringent real-time requirements of rail transit scenarios.

[0027] Third, this invention proposes a multi-scale enhancement encoder to fully mine multi-modal complementary information.

[0028] The enhanced encoder designed in this invention includes a multi-scale enhancement module and a spatial information recovery module. It can perform cross-scale interaction and global context modeling on backbone features of different resolutions, generating semantically rich global features. The spatial information recovery module adopts a multi-level upsampling-fusion structure, progressively restoring the global features to high resolution and fusing them step-by-step with the original multi-scale features, ultimately outputting low, medium, and high-resolution enhanced features containing rich spatial details. This design allows multi-modal data such as visible light, infrared, and radar to fully complement each other at the feature level, significantly improving the robustness of recognition in degraded environments such as haze, nighttime, and occlusion.

[0029] Fourth, this application can weaken the ability to perceive small targets.

[0030] In the U-Net structure of the residual prediction network, this invention customizes differentiated conditional injection strategies for features of different resolutions. At deep low-resolution features, a cross-attention module is used to establish a global dependency between medium-resolution conditional features and decoding features to achieve context awareness. At shallow high-resolution features, an AdaIN (Adaptive Instance Normalization) module is introduced to perform pixel-level style modulation on decoding features using high-resolution conditional features, accurately restoring high-frequency details such as the edges and textures of the target. This enables the model to achieve sub-pixel accuracy in boundary localization and category discrimination of small targets, making it particularly suitable for the recognition of extremely difficult samples such as long-distance pedestrians and small obstacles in rail transit.

[0031] Fifth, this invention employs a pseudo-truth value generation and refinement strategy.

[0032] This invention addresses the problems of high cost and poor consistency in manual annotation of small targets by proposing a complete pseudo-ground value generation and refinement process. First, multiple advanced detection models are used to form a model pool to generate candidate detection results. Second, a task-aware scoring function (weighted recall, precision, small target AP, etc.) is designed for small targets to automatically select the optimal pseudo-labels. Finally, a pre-trained refinement module decomposes the pseudo-labels, restores and reconstructs high-frequency details, and generates high-quality pseudo-ground values ​​aligned with the real data manifold. This method obtains massive amounts of reliable supervision signals without additional manual annotation, enabling rapid iterative deployment of the model in new scenarios and greatly improving the method's practicality and scalability.

[0033] In an optional embodiment of the present invention, a joint training method is used to train the first residual diffusion model and the second residual diffusion model. Pre-constructed samples are used as training data, and pre-generated pseudo-true values ​​are used as sample labels. The initial first and second residual diffusion models are trained to obtain a trained first residual diffusion model and a trained second residual diffusion model, including: Multimodal data was acquired and registered, cropped, and normalized. The preprocessed multimodal data was used as samples to construct a training dataset. It should be noted that degradation-clarification data pairs of various degradation types were synthesized, and the samples were degradation-clarification data pairs. The degradation included various degradation types, such as common degradations in rail transit such as haze, raindrops, motion blur, low light, and sensor noise. The samples were normalized to construct the training dataset. Obtain the pseudo-true values ​​of samples in the training dataset and use them as the labels for the samples; Input a portion of the samples from the training dataset into the first... The first and second residual diffusion models to be trained are then used to obtain the second residual diffusion model. The prediction results output during this training process; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the samples from the first and second residual diffusion models to be trained are used to calculate the loss and serve as the basis for the training. Loss during each training session; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the first and second residual diffusion models to be trained are obtained. The first and second residual diffusion models are to be trained. This process is repeated until the number of training iterations or the degree of convergence meets the preset conditions, at which point the trained first and second residual diffusion models are obtained.

[0034] Furthermore, in this embodiment, obtaining the pseudo-true values ​​of samples in the training dataset as the labels of the samples includes: A model pool consisting of multiple object detection models is used to detect samples in the training dataset, and each object detection model outputs the corresponding detection result; Based on the characteristics of the small target detection task, an evaluation index is selected, and the detection results output by each target detection model are weighted and scored. The detection result with the highest score is selected as the preliminary false true value. A pre-trained refining module is used to decompose and enhance the initial pseudo-true values ​​to obtain high-quality pseudo-true values, which are then used as labels for the samples.

[0035] It should be noted that, due to the high cost and error-prone nature of manual annotation for small objects, this embodiment employs a pseudo-ground value generation strategy. For the input degenerate sample pairs, a model pool consisting of multiple object detection models is used for detection. The object detection models can be YOLOX, NanoDet, PP-YOLOE, etc. The detection results output by each object detection model include bounding boxes and categories. Next, a task-aware scoring function is designed. Based on the characteristics of the small object detection task, evaluation metrics are selected, such as recall, precision, F1 score, and small object mean precision (mAP). The results output by each object detection model are weighted and scored, and the result with the highest score is selected as the initial pseudo-label. Next, a pre-trained refining module, such as DetBooster, is used to decompose and enhance the initial pseudo-labels, generating high-quality small object pseudo-ground values ​​as the labels for the samples.

[0036] Furthermore, in this embodiment, the loss The expression is: ; ; ; in, This indicates the loss in super-resolution reconstruction. The weights represent the super-resolution reconstruction loss. This indicates a loss in the detection task. The weights representing the loss of the detection task. Indicates consistency loss. The weights representing the consistency loss Indicates pixel-level loss. The weights represent the pixel-level loss. Indicates perceived loss. The weights representing the perceived loss Labels representing samples, This indicates the prediction result for the sample.

[0037] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0038] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0039] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for identifying small targets in rail transit based on a two-stage residual diffusion model, characterized in that, include: Acquire multimodal data to be identified; wherein, the multimodal data to be identified includes small targets; The multimodal data to be identified is input into a trained first residual diffusion model for processing to predict key region features and reconstruct the multimodal data to be identified. The reconstructed multimodal data and the multimodal data to be identified are then input into a trained second residual diffusion model for processing to obtain the identification result of the small target. The first residual diffusion model and the second residual diffusion model are trained by joint training. The pre-constructed samples are used as training data and the pre-generated pseudo-true values ​​are used as the labels of the samples. The initial first residual diffusion model and the second residual diffusion model are trained to obtain the trained first residual diffusion model and the trained second residual diffusion model.

2. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 1, characterized in that, The trained first residual diffusion model and the trained second residual diffusion model have the same architecture, both including: a backbone network, an augmented encoder, a residual diffusion model, and a decoder; wherein, The backbone network in the first residual diffusion model is used to extract multi-scale features from the multimodal data to be identified, thereby obtaining the first multi-scale features; The enhancement encoder in the first residual diffusion model is used to perform multi-scale compression and global feature enhancement on the first multi-scale feature, and to fuse the enhanced global feature with the first multi-scale feature to obtain the first enhanced feature; The residual diffusion model in the first residual diffusion model is used to predict the first enhanced feature to obtain the first structural feature; The decoder in the first residual diffusion model is used to reconstruct the first structural feature to obtain the reconstructed multimodal data to be identified; The backbone network in the second residual diffusion model is used to extract multi-scale features from the spliced ​​data of the multimodal data to be identified and the reconstructed multimodal data to be identified, so as to obtain the second multi-scale features. The enhanced encoder in the second residual diffusion model is used to perform multi-scale compression and global feature enhancement on the second multi-scale feature, and then fuses the enhanced global feature with the second multi-scale feature to obtain the second enhanced feature; The residual diffusion model in the second residual diffusion model is used to predict the second enhanced feature to obtain the second structural feature; In the second residual diffusion model, the decoder is used to reconstruct the second structural feature to obtain the recognition result of the small target.

3. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 2, characterized in that, The enhanced encoder includes a multi-scale enhancement module and a spatial information recovery module; wherein... The enhancement module is used to fuse and enhance the multi-scale features to obtain enhanced global features; The spatial information recovery module is used to fuse the enhanced global features and the multi-scale features to obtain enhanced features.

4. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 3, characterized in that, The spatial information recovery module includes a first upsampling module, a second upsampling module, a third upsampling module, a first downsampling module, a convolution module, a first fusion module, and a second fusion module; wherein, The enhanced global features are downsampled using the first downsampling module to obtain the first feature; The convolution module is used to perform a convolution operation on the feature concatenated from the first feature and the third scale feature to obtain the second feature; The first upsampling module upsamples the second feature to obtain the third feature; the third feature, the second scale feature, and the enhanced global feature are concatenated to obtain the fourth feature; The fourth feature is fused using the first fusion module to obtain the fifth feature; The fifth feature is upsampled using the second upsampling module to obtain the sixth feature; The enhanced global features are upsampled using the third upsampling module to obtain the seventh feature; The seventh feature, the first scale feature, and the sixth feature are concatenated to obtain the eighth feature; the eighth feature is fused using the second fusion module to obtain the ninth feature; the second feature, the fifth feature, and the ninth feature are concatenated to form an enhanced feature; wherein the second feature is used as a low-resolution feature, the fifth feature is used as a medium-resolution feature, and the ninth feature is used as a high-resolution feature.

5. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 4, characterized in that, The first fusion module and the second fusion module have the same structure, both including a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer; wherein, The first convolutional layer is used to perform a convolution operation on the input features to obtain the first convolutional features; The second convolutional layer is used to perform a convolution operation on the input features to obtain the second convolutional features; The third convolutional layer is used to perform a convolution operation on the first convolutional feature to obtain the third convolutional feature; The fourth convolutional layer is used to perform a convolution operation on the third convolutional layer to obtain the fourth convolutional feature; The fifth convolutional layer is used to perform a convolution operation on the first convolutional feature to obtain the fifth convolutional feature; The fifth convolutional feature is added element-wise to the fourth convolutional feature to obtain the sixth convolutional feature; The sixth convolutional feature is added element by element to the second convolutional feature to obtain the output feature.

6. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 2, characterized in that, The residual diffusion model includes a conditional mapping network and a residual prediction network, wherein the residual prediction network adopts a U-Net structure; wherein... The conditional mapping network is used to map the enhanced features into preliminary estimated features; Using multi-scale resolution features as conditions, the residual prediction network is used to denoise the noise of random sampling to predict clean residuals. The clean residuals are then intersected with the preliminary estimated features to obtain the output of the residual diffusion model.

7. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 6, characterized in that, The residual prediction network includes an upsampling module, a bottleneck module, and a downsampling module. The upsampling module includes a first residual module, a second residual module, and a third residual module, each of which includes two residual blocks. The bottleneck module includes a third fusion module. The downsampling module includes a fourth residual module, a fifth residual module, and a sixth residual module, each of which includes two residual modules. The fourth residual module further includes a cross-attention module, and the sixth residual module further includes an AdaIN module. The first residual module is used to process the randomly sampled noise and time step embedding vector to obtain the first residual feature; The second residual module is used to process the first residual feature to obtain the second residual feature; The third residual module is used to process the first residual feature to obtain the third residual feature; The bottleneck module is used to fuse the third residual feature with the low-resolution feature to obtain the fused feature; The fused features are processed using the fourth residual module to obtain the fourth residual features. The cross-attention module is used to fuse the medium-resolution feature and the fourth residual feature to obtain the second fused feature; the second fused feature and the second residual feature are added element-wise to obtain the third fused feature; The fifth residual module is used to process the third fusion feature to obtain the fifth residual feature. The fifth residual feature is then added element-wise to the first residual feature to obtain the fourth fusion feature. The sixth residual module is used to process the fourth fusion feature to obtain the sixth residual feature; Based on the high-resolution features, the AdaIN module is used to fine-tune the sixth residual feature to obtain the output of the residual prediction network.

8. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 1, characterized in that, A joint training approach is used to train the first and second residual diffusion models. Pre-constructed samples are used as training data, and pre-generated pseudo-true values ​​are used as sample labels. This process trains the initial first and second residual diffusion models, resulting in trained first and second residual diffusion models, including: Acquire multimodal data, and perform registration, cropping and normalization on the multimodal data. Use the preprocessed multimodal data as samples to construct a training dataset. Obtain the pseudo-true values ​​of the samples in the training dataset and use them as the labels of the samples; Input a portion of the samples from the training dataset into the first... The first and second residual diffusion models to be trained are then used to obtain the second residual diffusion model. The prediction results output during this training process; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the samples from the first and second residual diffusion models to be trained are used to calculate the loss and serve as the basis for the training. Loss during each training session; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the first and second residual diffusion models to be trained are obtained. The first and second residual diffusion models are to be trained. This process is repeated until the number of training iterations or the degree of convergence meets the preset conditions, at which point the trained first and second residual diffusion models are obtained.

9. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 8, characterized in that, Obtain the pseudo-true values ​​of samples in the training dataset as the labels of the samples, including: A model pool consisting of multiple object detection models is used to detect samples in the training dataset, and each object detection model outputs a corresponding detection result; Based on the characteristics of the small target detection task, an evaluation index is selected, and the detection results output by each target detection model are weighted and scored. The detection result with the highest score is selected as the preliminary false true value. The preliminary pseudo-true values ​​are decomposed and enhanced using a pre-trained refining module to obtain high-quality pseudo-true values, which are then used as labels for the samples.

10. The method for identifying small targets in rail transit based on a two-stage residual diffusion model according to claim 8, characterized in that, The loss The expression is: ； ； ； in, This indicates the loss in super-resolution reconstruction. The weights represent the super-resolution reconstruction loss. This indicates a loss in the detection task. The weights representing the loss of the detection task. Indicates consistency loss. The weights representing the consistency loss Indicates pixel-level loss. The weights represent the pixel-level loss. Indicates perceived loss. The weights representing the perceived loss Labels representing samples, This indicates the prediction result for the sample.

Images