Intelligent traffic target detection method and system based on deep sparse representation

Abstract

The present application relates to the technical field of image processing, and particularly relates to an intelligent traffic target detection method and system based on deep sparse representation; it comprises an adaptive compressive sensing module, a key area of a road scene is recognized based on saliency detection, then adaptive sampling oriented by a task is realized by updating a measurement matrix, so that high-dimensional image data is mapped into low-dimensional compressed representation; a light-weight network module reduces model complexity based on deep separable convolution and pruning technology, then knowledge distillation technology is introduced to transfer the detection ability of a teacher model to a student model to maintain high detection precision; an end-to-end optimization module drives the collaborative learning of a sampling matrix and a detection network through a joint loss function, realizing the whole-process optimization from data acquisition to target detection. Through the organic combination of the three modules, the proposed framework realizes effective target detection based on the compressed measurement domain, significantly reducing the computational complexity while ensuring the detection precision.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an intelligent traffic target detection method and system based on deep sparse representation. Background Technology

[0002] With the rapid development of autonomous driving technology, real-time and accurate target detection has become a core technology for ensuring driving safety. Most existing target detection algorithms are based on feature extraction and detection from high-resolution raw images, resulting in high computational complexity and poor real-time performance, making it difficult to meet the low latency requirements of autonomous driving systems. At the same time, it puts enormous data transmission and storage pressure on edge computing devices.

[0003] Existing solutions mainly focus on two directions: data compression and computational optimization. Data compression methods mostly follow the traditional paradigm of "compress first, then process," requiring complete decoding before detection, and the compression and detection modules are relatively independent. Computational optimization methods mostly focus on lightweight design at the model level, failing to fundamentally solve the burden of acquisition, transmission, and storage brought about by high-resolution image data.

[0004] The emergence of compressed sensing theory has provided a new solution to this problem. Based on signal sparsity, it achieves "sampling as compression," significantly reducing data overhead. However, traditional compressed sensing methods suffer from complex and time-consuming reconstruction algorithms. Although deep learning technology has improved the reconstruction efficiency of compressed sensing, existing deep learning-based compressed sensing detection methods still require image reconstruction before detection, which can easily lead to information distortion and error accumulation, affecting detection accuracy. While compressed learning theory has proven that compressed measurements can be directly used for inference, existing compressed learning models suffer from high computational complexity, use fixed sampling matrices, and lack dynamic scene adaptation capabilities, making them unsuitable for intelligent transportation target detection tasks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent traffic target detection method based on deep sparse representation, which eliminates the need for image reconstruction and directly performs target detection in the compressed measurement domain, balancing detection accuracy and computational efficiency.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides an intelligent traffic target detection method based on deep sparse representation, comprising the following steps: S1. Construct an adaptive compressed sensing module and process the image based on the adaptive compressed sensing module; The process includes: S11. Identify key regions of the image based on saliency detection, and dynamically allocate sampling density to the key regions and non-key regions according to a three-level hierarchical sampling strategy; S12. The image is mapped to a low-dimensional compressed representation using a learnable joint multichannel sampling matrix; S13. Optimize the feature discriminability of the compressed representation by maximizing the inter-class difference loss function, and selectively update the joint multi-channel sampling matrix based on parameter importance evaluation; S2. The obtained low-dimensional compressed representation is input into a lightweight compressed domain detection network for target recognition and localization. The lightweight compressed domain detection network adopts depthwise separable convolution and combines structured pruning and weight pruning strategies to simplify the network structure. S3. A knowledge distillation strategy is introduced for joint training. The complete YOLO network pre-trained based on the original scene image is used as the teacher model. Through the dual constraints of feature distillation and detection distillation, the image domain detection knowledge is transferred to the lightweight student model. S4. Construct a multi-objective joint loss function, integrating the basic detection loss, knowledge distillation loss, and inter-class difference maximization loss generated in the above steps. Perform end-to-end collaborative optimization of the sampling matrix and detection network through a progressive training strategy. Finally, the optimized lightweight compressed domain detection network directly completes target detection in the compressed measurement domain and outputs the detection results.

[0007] Preferably, in step S11, the saliency detection uses a method combining gradient and color contrast to calculate and obtain a saliency map; The saliency map, after being normalized and Gaussian filtered, is used to guide adaptive sampling.

[0008] Specifically, the image gradient features are calculated using the Scharr operator, and combined with the color contrast features of the LAB color space, a weighted saliency map is generated. in, , As a weighting factor, Indicates the gradient magnitude. Indicates color ratio characteristics.

[0009] Preferably, in step S11, the three-level hierarchical sampling strategy divides the scene image into multiple fixed-size sub-regions, performs corresponding block processing, calculates the saliency of each block through average pooling, and determines the hierarchical threshold using the statistical quantile method, dividing the scene image blocks into three sampling levels: high density, medium density, and low density.

[0010] Preferably, in step S12, the multi-channel sampling matrix is: in, This indicates three color channels. There are three sampling density levels. Indicates the channel index. Indicates density level; Measurement Dimensions Corresponding density level: in, The total dimension of the image patch. Based on the compression sensing ratio parameter , , These represent the measurement dimensions corresponding to the high, medium, and low sampling density levels, respectively. min() indicates taking the minimum value, and max() indicates taking the maximum value. During the sampling process, the sub-region is separated into three color channel components: R, G, and B. Then based on the sampling level Match the sampling matrix corresponding to channel c and density identifier r. This enables the following compressed sensing sampling: in, The corresponding compression measurement results.

[0011] Preferably, in step S13, the method for obtaining the parameter importance of the multi-channel sampling matrix is ​​as follows: ; ; ; in, For the multi-channel sampling matrix, the first The overall importance score of each parameter. , The importance of gradient and parameter magnitude are respectively. For the sampling matrix, the first Each parameter value; Total loss function For parameters The gradient represents the degree of influence of parameter changes on the loss; It is the sum of the absolute values ​​of the features in the measurement domain, reflecting the overall activation level of the current features; Based on the detection of loss, To maximize the loss for inter-class differences, , , These are feature contribution weights, task weights, and discriminative weights, respectively. Here is the numerical stability constant. These are the corresponding weighting coefficients.

[0012] Preferably, in step S13, the low importance parameter is updated based on the following formula, and then the joint multi-channel sampling matrix is ​​updated: in, For elements in the sampling matrix Current parameter value, To update the obtained parameter values, The learning rate controls the step size for updating parameters. This is the gradient of the loss function with respect to this parameter.

[0013] Preferably, in step S13, the update mechanism of the joint multi-channel sampling matrix is ​​constructed based on key indicators such as feature variance, feature amplitude, task effectiveness, and discrimination effectiveness to evaluate the quality of measurement domain information.

[0014] The specific methods for evaluating the quality of information in the measurement domain are as follows: ; in, To compress the feature variance in the domain, representing feature diversity and information richness; For feature vectors Norm, a measure of feature activation strength; These are the weighting coefficients for each component; For mission losses, To maximize the loss for differences between classes.

[0015] Preferably, in step S2, the depthwise separable convolution includes depthwise convolution and pointwise convolution, which are responsible for spatial local feature extraction and cross-channel feature fusion, respectively.

[0016] Preferably, in step S2, the structured pruning strategy is based on gradient sensitivity. Weight magnitude Characteristic response intensity and task relevance Construct the following comprehensive weighted score : in, For the corresponding weighting coefficients, These are the weight parameters of the sampling matrix. The gradient vector of the loss function. Let be the partial derivative with respect to the weight parameters; Weighted pruning strategy, based on the set pruning threshold Remove weights with small values ​​to obtain sparse connections: in, For position Pruning mask, For the corresponding weight parameters; The specific pruning ratio will vary over time. The changes are shown in the following formula: in, and These represent the current training round and the total training round, respectively. The target pruning ratio.

[0017] On the other hand, the present invention also provides a target detection system based on deep sparse representation applicable to the above-described method, comprising: The compressed sensing module is used to build an adaptive compressed sensing module to perform saliency detection on scene images to identify key regions. It dynamically adjusts the sampling density based on a three-level hierarchical sampling strategy, maps high-dimensional scene images to low-dimensional compressed representations through a learnable joint multi-channel sampling matrix, enhances the discriminability of targets in the compressed domain by maximizing the inter-class difference loss function, and achieves selective updating of the sampling matrix based on parameter importance evaluation. The lightweight compressed domain detection module is used to receive the low-dimensional compressed representation, construct a lightweight network using depthwise separable convolution combined with structured pruning and weight pruning strategies, extract features in the compressed domain and output preliminary detection results. The cross-domain knowledge distillation module, based on the knowledge distillation strategy, is used to transfer the detection knowledge learned by the teacher network in the original pixel domain to the lightweight compressed domain detection module during the training phase through the dual constraints of feature distillation and detection distillation. The end-to-end collaborative optimization module is used to construct a multi-objective joint loss function that includes basic detection loss, knowledge distillation loss, and inter-class difference maximization loss. It performs collaborative optimization on the adaptive compressed sensing module and the lightweight compressed domain detection module through a progressive training strategy, and finally completes target detection directly in the compressed measurement domain.

[0018] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the target scene detection method based on deep sparse representation.

[0019] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the target scene detection method based on deep sparse representation.

[0020] Compared with existing technologies, this invention provides an intelligent traffic target detection method based on deep sparse representation, which has the following advantages: 1. By performing target detection directly in the compressed sensing measurement domain, the image reconstruction step in the traditional "compression-reconstruction-detection" process is eliminated. This avoids information distortion and error accumulation in the reconstruction process and significantly reduces the time overhead of data processing, effectively meeting the low latency requirements of autonomous driving scenarios.

[0021] 2. The adaptive compressed sensing module based on saliency detection dynamically adjusts the sampling strategy based on the key regions of the detection task, and extracts saliency information in a targeted manner through a learnable sampling matrix. While reducing the data dimensionality, it maximizes the retention of high discriminative features of targets (such as vehicles and pedestrians), thereby improving the effective utilization rate of compressed domain data.

[0022] 3. By combining knowledge distillation technology to construct a lightweight model, and leveraging the image domain detection knowledge transfer of the teacher model, the shortcomings of the lightweight network's representation ability are compensated. While significantly reducing the number of model parameters and computational complexity, the accuracy of target detection is guaranteed, and it is adapted to the resource constraints of edge computing devices.

[0023] 4. The combination of adaptive sampling strategy and direct detection in compressed domain can better cope with information fluctuations in complex traffic scenarios (such as dense targets, occlusion, and day-night cycles), and improve the robustness and reliability of the detection system in actual autonomous driving environments.

[0024] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of a target scene detection method based on deep sparse representation. Figure 2 This is a diagram of the compressed sensing sampling process; Figure 3 Flowchart for updating the multi-channel sampling matrix; Figure 4 A diagram of a lightweight compressed domain detection network architecture; Figure 5 A diagram illustrating the principle of the knowledge distillation strategy; Figure 6 The figure shows a comparison between the proposed method and the reconstruction-based detection method. Figure 7 The figure shows a comparison between the proposed method and the non-reconstruction-based detection method. Figure 8 The figure shows a comparison between the proposed method and mainstream object detection methods. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0029] Example 1: The target scenarios applicable to this invention include roads, waterways, and air routes, such as... Figure 1 As shown, this embodiment provides a road target detection method based on deep sparse representation, including the following steps: S1. Construct an adaptive compressed sensing module to perform saliency detection on road scene images to identify key regions. Dynamically adjust the sampling density based on a three-level hierarchical sampling strategy. Map the high-dimensional road scene image to a low-dimensional compressed representation through a learnable joint multi-channel sampling matrix. Enhance the discriminability of the target in the compressed domain by maximizing the inter-class difference loss function. And achieve selective updating of the sampling matrix based on parameter importance evaluation. like Figure 2 As shown, during the compressed sampling process, the system can adaptively focus on important regions of the image, reducing the amount of data while preserving as much discriminative information as possible that is beneficial for target detection. This adaptive compressed sensing sampling module mainly consists of four sub-modules: a saliency detection sub-module, which detects salient regions in the road scene to provide reliable spatial prior information for adaptive sampling; The adaptive sampling submodule dynamically adjusts the sampling density level of the corresponding region based on the obtained saliency map, achieving efficient dimensionality reduction mapping from the original image to the compressed measurement domain. The inter-class difference maximization loss function enhances the separability of various targets within the compressed measurement domain, ensuring that the compression process effectively preserves the key discriminative information required for the detection task. The sampling matrix parameter update mechanism selectively updates the measurement matrix parameters using multi-dimensional evaluation results, prioritizing the retention of key parameters that contribute significantly to the detection task. Through the collaborative optimization of these four submodules, the adaptive compressed sensing module achieves deep coupling between the measurement process and the detection task, providing a compressed domain feature representation with both low dimensionality and high discriminative power for subsequent lightweight detection networks while ensuring measurement efficiency.

[0030] Preferably, the saliency detection submodule performs saliency modeling on the input image to identify key regions that significantly impact the target detection task. It employs a fast saliency detection method combining gradient and color contrast to calculate the saliency map. Specifically, the saliency detection module utilizes the Scharr operator to calculate image information gradient features to obtain relevant information characterizing key structures such as edges and contours. Simultaneously, it combines this with color contrast features from the LAB color space, which effectively characterizes the discriminative power of the target against the background, to generate a weighted saliency map as follows. : in, , As a weighting factor, Indicates the gradient magnitude. This represents the color contrast characteristics. The generated saliency map, after normalization and Gaussian filtering, effectively suppresses detail noise while enhancing structural clarity, providing a solid foundation for subsequent adaptive sampling.

[0031] Preferably, the adaptive sampling submodule, based on the spatial distribution characteristics of the obtained saliency map, introduces a three-level hierarchical sampling strategy. This strategy divides the image into multiple fixed-size sub-regions, performs corresponding block processing, and calculates the saliency of each block using average pooling. The three-level hierarchical strategy uses the statistical quantile method to determine the hierarchical thresholds, using the 70% and 30% quantiles of the saliency block distribution as thresholds to divide image blocks into three sampling levels. These correspond to low-density, medium-density, and high-density sampling strategies, respectively. , These represent the row and column indices of the sub-block in the image, respectively. To meet the measurement requirements of different sampling levels, multiple sampling matrices with channel separation and density stratification are designed. Compressed sensing sampling matrices at different density levels are constructed for the three color channels, thus forming the following joint multi-channel sampling matrix. ,in This indicates three color channels. There are three sampling density levels. Indicates the channel index. Indicates density level label, measurement dimension Corresponding density level: in, The total dimension of the image patch. Based on the compression sensing ratio parameter , , These represent the measurement dimensions corresponding to the high, medium, and low sampling density levels, respectively. `min()` represents the minimum value, and `max()` represents the maximum value. The sampling rate is defined as follows: In the actual sampling process, the image is decomposed into multiple sub-blocks, and each sub-block can be separated into three color channel components: R, G, and B. Based on sampling level Matching and Channels and density label corresponding sampling matrix This enables the following compressed sensing sampling: The generated measurement results constitute a compressed sensing feature map. During this process, highly significant sub-blocks are sampled using a high-density sampling matrix. To effectively preserve key visual details such as edges and textures, the basic sub-blocks use a reference sampling matrix. To balance information fidelity and compression efficiency, low-saliency sub-blocks utilize a low-density sampling matrix. This achieves efficient compression of the background region. Through the aforementioned sub-block and channel-by-channel adaptive sampling matrix selection mechanism, spatial differentiation of measurement resources within a single image can be effectively realized, thereby ensuring that key regions can obtain sufficient measurement resources.

[0032] Based on the design concept of an "optimizable joint multi-channel sampling matrix," the measurement process of traditional compressed sensing can be mapped to a trainable convolutional operation, thereby achieving end-to-end optimization of the compressed sensing sampling process. The fixed measurement matrix relied upon by traditional compressed sensing cannot adaptively adjust based on the dynamic characteristics of the signal, making it difficult to capture the essential spatiotemporal characteristics of the data. Deep learning models, on the other hand, automatically identify and utilize the inherent structural characteristics of the data through a learnable measurement process to significantly improve sampling performance. They parameterize the measurement process as convolutional layers; the local connectivity and weight sharing characteristics of convolutional operations can greatly reduce the parameter size while effectively capturing local spatial correlations in the image, thus achieving seamless end-to-end integration between the sampling process and the depth detection model. Specifically, the sampling matrix is ​​regarded as the weight kernel of a groupable convolution, grouped and encoded according to color channels, with each group corresponding to a subset of the measurement matrix for a single channel; each image employs a non-overlapping block sampling strategy, and the number of output channels depends on the measurement dimension. Dynamic settings enable dimensionality reduction mapping from raw data to a compressed measurement domain. This design embeds a learnable sampling matrix into a deep model, thereby achieving collaborative optimization of environmental perception and target detection.

[0033] Preferably, the inter-class difference maximization loss function is used. It is well known that signal dimensionality reduction inevitably results in information loss; therefore, it is necessary to ensure that the information obtained in the compressed domain has optimal discriminative effectiveness for target detection. Based on the criterion of maximizing target scene discriminability, the following inter-class difference maximization loss function based on Euclidean distance in the feature space is constructed: in, The intra-class distance loss is designed to minimize the distance between similar target samples in the feature space of the measurement domain, drive the clustering of similar feature vectors, and thus enhance the consistency of feature representation. This represents the inter-class distance loss, which aims to maximize the distance between target samples of different classes, thereby improving the model's ability to distinguish class differences. The minimum inter-class distance threshold is used to maintain the minimum discriminative margin between compressed features of each class. The constructed loss function enhances the model's sensitivity to inter-class differences by imposing global structural constraints on the measurement domain features, thereby improving the target discrimination capability of the compressed measurement domain. This allows the compressed sensing sampling matrix to effectively capture feature differences between targets, thus enhancing the target detection performance in the measurement domain.

[0034] Preferably, the sampling matrix parameter update mechanism is used because the quality of the sampling matrix directly affects the extraction of compressed domain features and the subsequent target detection performance. To improve the effectiveness of compressed sensing feature extraction, a mechanism is constructed as follows: Figure 3 The diagram illustrates a learnable sampling matrix parameter update mechanism based on task performance and feature quality feedback. This mechanism borrows from incremental learning, progressively optimizing and updating the learnable sampling matrix parameters with new data. This maintains the correlation between scenes while fully utilizing historical experience. By quantifying the contribution of each parameter to improving target detection performance, the mechanism prioritizes the retention of key parameters, effectively preventing the loss of important information.

[0035] During model training, the sampling matrix is ​​analyzed. The importance of each element is determined, and the sampling matrix parameters are selectively optimized based on this importance. The overall importance score of the sampling matrix parameters is as follows: in, For the sampling matrix, the first The overall importance score of each parameter. , The importance of gradient and parameter magnitude are respectively. For the sampling matrix, the first Each parameter value. Total loss function For parameters The gradient represents the degree to which parameter changes affect the loss. It is the sum of the absolute values ​​of the features in the measurement domain, reflecting the overall activation level of the current features. Based on the detection of loss, To maximize the loss for inter-class differences, , , These are feature contribution weights, task weights, and discriminative weights, respectively. Here is the numerical stability constant. These are the corresponding weighting coefficients.

[0036] During the sampling matrix update process, performance metrics such as task performance, feature quality, and inter-class discriminability are dynamically monitored, and parameters are adaptively adjusted based on metric changes. When a decrease in these performance metrics compared to historical benchmarks is detected, the sampling matrix parameter update mechanism is automatically triggered. The update process employs an importance-driven strategy: calculating the comprehensive importance score of each parameter position, identifying high-importance parameters requiring protection based on a set importance threshold, and updating low-importance parameters based on the following formula: in, For elements in the sampling matrix Current parameter value, To update the obtained parameter values, The learning rate controls the step size for updating parameters. This is the gradient of the loss function with respect to this parameter.

[0037] The aforementioned selective update mechanism effectively avoids the loss of key information that may result from global updates of the sampling matrix, and optimizes relevant parameters based on task performance feedback, thereby ensuring the integrity of semantic information in the compressed sensing domain. The update mechanism is constructed based on key indicators such as feature variance, feature magnitude, task effectiveness, and discriminative effectiveness to evaluate the quality of information in the measurement domain. in, To compress the feature variance in the domain, representing feature diversity and information richness; For feature vectors Norm, a measure of feature activation strength; denoted as the weight coefficients for each component. The proposed parameter update method based on importance assessment can effectively maintain the stability of key parameters while optimizing the sampling matrix to adapt to the data distribution characteristics, providing a more discriminative compressed domain information representation for subsequent target detection.

[0038] S2. Construct a lightweight compressed domain detection network, employing depthwise separable convolutions and combining structured pruning and weight pruning strategies to simplify the network structure; To efficiently process compressed measurement domain data, reduce network complexity, and ensure target detection accuracy, a lightweight network architecture based on CSPDarkNet is constructed, such as... Figure 4 As shown.

[0039] The network architecture adopts an end-to-end design approach. The backbone front-end integrates an adaptive compressed sensing module, which uses a trainable multi-channel sampling matrix to sample and process images, eliminating the need for reconstruction and directly detecting targets in the measurement domain. A parameter allocation strategy is employed, using a progressive channel reduction method to reduce computational complexity. Channels at each level are appropriately compressed to avoid feature representation degradation. Furthermore, a depthwise separable convolution module replaces the traditional standard convolution operation, decomposing it into two stages: depthwise convolution and pointwise convolution. These stages are responsible for spatial local feature extraction and cross-channel feature fusion, respectively, thereby reducing computational complexity and the number of parameters while maintaining receptive field and feature extraction capabilities. Depthwise convolution can be represented as: in, For the first Each channel is located in The depthwise convolution output, and These represent the vertical and horizontal indices of the convolution kernel, respectively. For the input feature map, the first Pixel values ​​of the channel, For depthwise convolutional kernel weights, This represents the kernel size.

[0040] Point convolution can be represented as: in, Indicates the first Each output channel is in position The final output, For point convolution weights, This is the number of input channels.

[0041] To further reduce model complexity, structured pruning based on importance assessment and weight pruning strategies are introduced. The structured pruning strategy is based on gradient sensitivity. Weight magnitude Characteristic response intensity and task relevance Construct the following comprehensive weighted score To evaluate the impact of each convolutional channel on detection performance: in, For the corresponding weighting coefficients, These are the weight parameters of the sampling matrix. The gradient vector of the loss function. Let be the partial derivative with respect to the weight parameters. Based on the importance assessment results, high-importance channels are retained, while low-importance channels are removed, thereby simplifying the network structure.

[0042] The weight pruning strategy uses fine-grained network weights as the pruning unit, directly selecting and removing convolutional kernel weights. This is based on a set pruning threshold. The following formula is constructed to remove weights with smaller magnitudes to obtain sparse connections: in, For position Pruning mask, For the corresponding weight parameters, This is the pruning threshold. Pruned branches with zero weights are no longer included in the calculation during forward propagation.

[0043] To ensure a smooth transition in model performance, a progressive pruning strategy is adopted, which involves gradually increasing the pruning intensity based on the training epochs, with the specific pruning ratio varying over time. The changes are shown in the following formula: in, and These represent the current training round and the total training round, respectively. The target pruning ratio is set. This strategy maintains model integrity in the early stages of training, gradually increases pruning intensity in the middle stages of training, and ensures that the structure compression reaches the preset ratio in the later stages of training, while maintaining the effective representation ability of key features.

[0044] S3. A knowledge distillation strategy is introduced, using a complete YOLO network pre-trained based on original road scene images as the teacher model. Through the dual constraints of feature distillation and detection distillation, image domain detection knowledge is transferred to a lightweight student model. While the aforementioned pruning strategies can effectively compress the model, the loss of feature domain information caused by compressed sensing sampling may still reduce the model's representational ability. To address this issue, based on the idea of ​​distillation models, a knowledge distillation training strategy is introduced to further reduce model complexity while improving the object detection performance of lightweight networks.

[0045] Specifically, the knowledge distillation strategy adopts a teacher-student distillation model framework. The teacher model is a complete YOLO network pre-trained on the original image, possessing strong feature representation capabilities. The student model is a lightweight network designed in this invention; however, due to unavoidable information loss in the measurement domain, its feature representation capabilities are relatively weak, and key details are easily lost. To address the cross-domain knowledge transfer problem, the proposed distillation model sets feature distillation points in the backbone network and multi-layer detector heads. Feature adapters align the feature dimensions of the teacher and student models. The distillation process is as follows: Figure 5 As shown. During training, by aligning the intermediate layer features of the student and teacher models, the student model can effectively inherit the teacher model's knowledge in feature extraction and object detection. The differences between the two can be addressed using... Norm metric, and based on this, the following feature distillation loss is constructed: in, Characteristic distillation loss, Number of distillation points and The student-teacher model is in the first... Characteristic representation of each distillation point, For the feature adaptation operator, dimension alignment is achieved through bilinear interpolation and 1×1 convolution. This distillation strategy enables the student model to learn the multi-scale spatial representation capabilities of the teacher model, thereby enhancing the model's representational ability and robustness.

[0046] To ensure consistent task awareness in the detection head output, a detection distillation strategy based on cosine similarity is introduced to align the orientation of the final output feature vector. The proposed detection distillation loss can be defined as follows: in, and ... in, For total distillation losses, Characteristic distillation loss, To detect distillation loss, The weights are used to dynamically adjust the weights of the two distillation branches during the training phase. Based on the above multi-level distillation strategy, the student model can effectively learn the feature representations of the teacher model while maintaining lightweight design, thereby improving the performance of object detection in the compressed domain.

[0047] S4. Construct a multi-objective joint loss function that integrates basic detection loss, knowledge distillation loss, and inter-class difference maximization loss. Through a progressive training strategy, achieve end-to-end collaborative optimization of the sampling matrix and the detection network, and directly complete target detection in the compressed measurement domain.

[0048] For the constructed lightweight compressed domain object detection model, a multi-objective joint optimization and progressive training strategy is adopted to achieve the collaborative optimization of the entire model. This training mechanism can simultaneously optimize the compressed sensing sampling matrix, network weights, and knowledge distillation process, thereby ensuring that each module performs collaborative optimization learning with the goal of maximizing the object detection probability.

[0049] The design of a multi-objective loss function is the core of achieving end-to-end optimization. By using various performance losses such as linearly weighted compressed sensing, object detection, and knowledge distillation, the following end-to-end training loss function can be obtained: in , , These are the basic detection loss, knowledge distillation loss, and inter-class difference maximization loss, respectively. , , These are the classification cross-entropy loss, confidence loss, and bounding box regression loss, respectively. , , Hyperparameters for controlling the loss of each component.

[0050] To address the complex optimization issues in the end-to-end training process, a progressive training strategy is adopted to ensure that the compressed sensing module and object detection network can stably implement hierarchical optimization. The training process is divided into a warm-up phase, a feature learning phase, and an end-to-end optimization phase. In the warm-up phase, the compressed sensing module and network backbone are initialized and trained using a low learning rate to optimize the compressed sensing sampling capability, ensuring that the model can extract effective features. In the feature learning phase, a knowledge distillation mechanism is introduced to guide the student model to learn intermediate feature representations from the teacher model, enhancing the feature representation capability of the lightweight model. In the end-to-end optimization phase, the compressed sensing module, lightweight detection network, and knowledge distillation module are jointly trained, and a unified multi-objective loss function is used for global optimization.

[0051] This experiment is based on the PyTorch deep learning framework, with an Intel Core™ i9-13900K processor, a clock speed of 3.00GHz, 32GB of memory, and an NVIDIA GeForce RTX 3090 GPU.

[0052] The evaluation metric used is mAP (Mean Average Precision), a commonly used evaluation metric in object detection models. mAP@0.5 represents the average precision when the IoU threshold is 0.5, and mAP@0.5:0.95 represents the average precision calculated with a step size of 0.05 when the IoU threshold ranges from 0.5 to 0.95. Param(M) represents the number of model parameters, and GFLOPs represents the floating-point computation cost of the model.

[0053] Traditional compressed sensing object detection methods employ a pipelined processing model of "sampling, reconstruction, and object detection." Since compressed sensing reconstruction is essentially an ill-conditioned inverse problem, the reconstruction stage inevitably leads to the loss of discriminative features relevant to the detection task, thus degrading the object detection performance based on the reconstructed image. Experiments selected ReconNet, ISTANET, and CSformer as compressed sensing reconstruction networks, and YOLOv8n as the detector, with a uniform compression ratio of 0.25.

[0054] Depend on Figure 6It is known that traditional "reconstruction before detection" methods introduce information distortion during the reconstruction stage, resulting in significant false positives or false negatives in small target vehicle detection. Specifically, ReconNet employs a block processing strategy, processing each image block independently, lacking inter-block information interaction. This leads to noticeable block artifacts at block boundaries in the reconstructed image, destroying target edge contour information and causing false negatives for distant vehicle targets. The iterative phase inherent in ISTA-NET struggles to adapt to different complex scenes, especially densely populated vehicle scenes, making it difficult for the reconstruction process to converge sufficiently, resulting in blurred boundaries between adjacent targets. Regarding the CSformer model, although it has a Transformer self-attention mechanism that can effectively process temporal information in parallel, it still struggles to fully compensate for the high-frequency information lost during the sampling process when processing compressed measurement data, leading to a decrease in the accuracy of small target localization. In contrast, the proposed method can effectively distinguish targets in densely populated and mutually occluded vehicle scenes, while significantly reducing false negatives for distant small-scale vehicles. This can be attributed to the fact that the proposed method does not require image reconstruction and directly detects targets based on the compressed measurement domain, thus effectively avoiding information distortion and error accumulation during the reconstruction process. Furthermore, the constructed adaptive compressed sensing module can dynamically adjust the sampling strategy based on scene saliency, thus retaining rich target discrimination information in key areas. In addition, the end-to-end optimization framework enables the sampling matrix and detection network to learn collaboratively, thereby achieving end-to-end optimization.

[0055] To further verify the effectiveness of the proposed method, the detection results of non-reconstruction detection methods CCN, SP-ILC, and the proposed method were compared with those of other methods. Figure 7 As shown in the figure, CCN has limited ability to distinguish adjacent targets in densely trafficked scenes and complex weather conditions. Although the SP-ILC method can detect the main targets, it exhibits significant false detections. The reasons for this phenomenon can be summarized as follows: CCN lacks a dedicated feature extraction design for the measurement domain, making it difficult to fully exploit the discriminative information in compressed data. Consequently, it struggles to separate adjacent targets in densely trafficked scenes, leading to missed detections and false detections. The SP-ILC method employs a multi-task learning framework to simultaneously optimize imaging, localization, and classification tasks. The competition among these tasks makes it difficult to simultaneously achieve both detection accuracy and localization accuracy in complex scenes. In contrast, the detection boxes generated by the proposed method show better consistency with the actual vehicle locations in complex scenes, and the localization results exhibit high stability across different scenarios. This is mainly attributed to the specific feature extraction module designed in the proposed method, which can effectively learn the inherent representation of compressed measurement data. Through an adaptive sampling mechanism and end-to-end optimization strategy, the sampling process and detection task are deeply coupled, avoiding competition between tasks and thus improving detection performance.

[0056] To comprehensively evaluate the detection performance of the proposed method, it can be compared with the following mainstream object detection algorithms: single-stage detectors RetinaNet, CenterNe, and YOLOv8n; two-stage detector Faster R-CNN; and RT-DETR based on the Transformer architecture. The detection results are as follows: Figure 8 As shown in the figure, RetinaNet is prone to target omissions and false detections. CenterNet's overall detection results are relatively stable, but its bounding box localization accuracy is insufficient in some scenarios. YOLOv8n performs well in most scenarios, but its detection accuracy decreases under low-light conditions at night. Faster R-CNN and RT-DETR both exhibit some degree of duplicate detection in dense target areas. This is because RetinaNet uses a fixed anchor box scale, making it difficult to adapt to vehicle targets of different sizes. Furthermore, in dense scenes, the anchor box mechanism is prone to causing an imbalance between positive and negative samples, thus affecting detection performance. CenterNet, on the other hand, indirectly regresses bounding boxes through the target center point. When the visual features of the target center area are not significant, it is prone to introducing cumulative and quantization errors, leading to a decrease in localization accuracy. YOLOv8n is susceptible to noise interference and reduced contrast in low-light scenes at night, and its feature extraction ability is weak. For Faster R-CNN, its region proposal network and detection network predict independently, and the same target may generate multiple high-confidence candidate boxes. However, non-maximum suppression with a fixed threshold is difficult to effectively remove redundant detections in dense scenes. Regarding RT-DETR, its decoder self-attention mechanism in dense scenes easily leads to multiple query vectors focusing on the same target, and it lacks an explicit post-processing suppression strategy, resulting in duplicate detections. In contrast, the proposed method can maintain stable detection output in various complex scenes, and still has good target separation performance under dense vehicle and occlusion conditions. This is because the proposed model is based on direct detection in the compressed sensing domain, allocates more sampling resources to key regions through an adaptive sampling mechanism, and avoids error accumulation problems through lightweight detection and end-to-end optimization strategies.

[0057] Based on the BDD100K dataset, the results of the four evaluation metrics obtained by the above detection methods based on reconstruction, non-reconstruction, and mainstream object detection methods are shown in Tables 1-3. The bolded part is the optimal value of the corresponding column, and the underline indicates the corresponding suboptimal value.

[0058] Table 1 Comparison with Reconstruction-Based Detection Methods As shown in Table 1, the reconstruction-based detection methods perform poorly in terms of mAP@0.5 and mAP@0.5:0.95, while also exhibiting generally high computational costs. This is because reconstruction-based methods rely on reconstruction results for subsequent detection, and their performance is highly limited by reconstruction quality. Furthermore, the reconstruction process inevitably introduces noise and information distortion, which accumulate and propagate throughout the detection process, thus reducing overall detection accuracy. In contrast, the proposed method outperforms the three reconstruction-based detection methods in both of the aforementioned metrics, while also having lower parameters and computational costs, achieving a balance between performance and computational efficiency. This can be attributed to the fact that the proposed method directly learns detection-related features from the measurement domain, avoiding the accumulation and propagation of reconstruction errors. Moreover, regarding computational costs, traditional methods significantly increase the computational burden due to the introduction of additional reconstruction modules, while the proposed method can complete detection inference without reconstruction, ensuring detection accuracy while effectively reducing computational overhead.

[0059] Table 2 shows the quantitative comparison results of non-reconstruction-based detection methods. It can be seen that the proposed model outperforms the comparative methods in four metrics: mAP@0.5, mAP@0.5:0.95, Param, and GFLOPs, achieving a better balance between accuracy and complexity. While CNN and SP-ILC do not require reconstruction, they are not specifically designed based on the characteristics of compressed measurement structures, thus easily leading to insufficient semantic extraction and background interference in complex scenes. Furthermore, these methods rely on increasing network size or enhancing convolutional structures to improve performance, resulting in a significant increase in the number of parameters and computational cost. In contrast, the adaptive sampling module built by the proposed method for the measurement domain can effectively capture target information, thereby achieving higher detection accuracy and computational efficiency without relying on large-scale network structures.

[0060] Table 2 Comparison with non-reconstruction-based detection methods Table 3 shows the quantitative comparison results with mainstream object detection methods. As can be seen from Table 3, the proposed method outperforms the comparison methods in both mAP@0.5 and mAP@0.5:0.95, while maintaining low levels in Param and GFLOPs, only slightly higher than YOLOv8n, significantly outperforming most two-stage and single-stage detection methods. Therefore, it achieves high detection accuracy while maintaining low model complexity. The performance differences mentioned above mainly stem from the different input information modeling methods and feature representation strategies of various methods. Mainstream object detection methods are usually based on detecting objects in the complete image domain, with network structure designs biased towards high-fidelity input, and feature extraction processes highly dependent on rich texture information and spatial details. In contrast, the proposed method directly detects objects in the compressed domain, with input being compressed measurement data. It focuses more on the compact expression of effective semantic information and prioritizes encoding target-related regions through an adaptive sampling mechanism, thus enabling it to maintain high target-background discrimination ability even under limited input information, thereby achieving higher detection accuracy and efficiency.

[0061] Table 3 Comparison with mainstream target detection methods To verify the effectiveness of each module in the proposed model, a progressive verification strategy was adopted. Based on the baseline model of direct detection in the compressed measurement domain, an adaptive sampling module, a learnable matrix, and a knowledge distillation module were added in sequence to evaluate the contribution of each module to the target detection task step by step. The results are shown in Table 4.

[0062] Table 4 Ablation Experiment Results As shown in the table, the introduction of each module has a positive impact on detection performance, showing an overall gradual upward trend. The baseline model has relatively limited detection performance when using only a fixed sampling matrix. After introducing the adaptive sampling module, the model's detection accuracy is significantly improved. Based on this, further combining it with a learnable sampling matrix further improves detection performance, indicating that the learnable measurement mechanism can effectively enhance feature representation capabilities. Simultaneously introducing adaptive sampling, a learnable matrix, and knowledge distillation modules, the model achieves high target detection accuracy, thus verifying the effectiveness of the proposed model. The above performance improvement stems from the synergistic optimization of each module: the adaptive sampling module can dynamically adjust the sampling strategy based on the spatial distribution of the target, making the sampling process focus more on information-dense areas, thereby reducing the interference of irrelevant background information on the detection task. The learnable sampling matrix replaces the fixed measurement matrix with a trainable structure, allowing the model to adaptively learn a compression mapping that better meets the requirements of target detection during end-to-end training, thereby further improving the discriminative power of extracted features. The knowledge distillation mechanism effectively transfers the detection experience gained by the teacher model to the student model through the dual constraints of feature distillation, which ensures the consistency of the intermediate layer feature representation, and detection distillation, which improves the reliability of the final prediction results, thereby further optimizing the overall detection performance.

[0063] To further verify the effectiveness of the proposed method in feature extraction, the constructed adaptive compressed sensing module was replaced with the feature extraction modules of the following models: ECA-Net, CoordinateAttention, and TripletAttention. The detection results are shown in Table 5. Table 5 Comparison of different feature extraction modules As shown in the table, the proposed feature extraction method has a high mAP value. ECA-Stem achieves efficient channel attention modeling based on local cross-channel interactions; however, it lacks attention to spatially salient regions and discriminative structures. Coordinate-Stem captures positional information and long-range dependencies through directional encoding to enhance the representation of spatial structures; however, it does not consider the semantic requirements of the detection scenario, thus limiting its ability to model key regions specifically. Triplet-Stem models cross-dimensional dependencies based on three branches: height, width, and channel, possessing strong spatial and semantic enhancement capabilities; however, it lacks explicit task guidance and its selection of discriminative regions is not targeted. In contrast, the proposed adaptive compressed sensing module, through saliency detection, adaptive sampling, inter-class difference enhancement, and collaborative optimization of the learnable matrix, enables the model to focus on detecting relevant key regions and effectively retain highly discriminative features, thereby achieving superior compressed domain object detection performance.

[0064] Example 2: This embodiment provides a road target detection system based on deep sparse representation, including: The compressed sensing module is an adaptive compressed sensing module that performs saliency detection on road scene images to identify key regions. It dynamically adjusts the sampling density based on a three-level hierarchical sampling strategy, maps high-dimensional road scene images to low-dimensional compressed representations through a learnable joint multi-channel sampling matrix, enhances the discriminability of targets in the compressed domain by maximizing the inter-class difference loss function, and achieves selective updating of the sampling matrix based on parameter importance evaluation. The lightweight detection module constructs a lightweight compressed domain detection network, employing depthwise separable convolutions and combining structured pruning and weight pruning strategies to simplify the network structure. The cross-domain knowledge distillation module introduces a knowledge distillation strategy. It uses a complete YOLO network pre-trained based on original road scene images as the teacher model. Through the dual constraints of feature distillation and detection distillation, it transfers image domain detection knowledge to a lightweight student model. The end-to-end collaborative optimization module constructs a multi-objective joint loss function, integrating basic detection loss, knowledge distillation loss, and inter-class difference maximization loss. Through a progressive training strategy, it achieves end-to-end collaborative optimization of the sampling matrix and the detection network, directly completing target detection in the compressed measurement domain.

[0065] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0066] Example 3: An electronic device is provided for running the aforementioned "target scene detection method based on deep sparse representation". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S4 of the method described in Embodiment 1, specifically including but not limited to: S1. Construct an adaptive compressed sensing module to perform saliency detection on road scene images to identify key regions. Dynamically adjust the sampling density based on a three-level hierarchical sampling strategy. Map the high-dimensional road scene image to a low-dimensional compressed representation through a learnable joint multi-channel sampling matrix. Enhance the discriminability of the target in the compressed domain by maximizing the inter-class difference loss function. And achieve selective updating of the sampling matrix based on parameter importance evaluation. S2. Construct a lightweight compressed domain detection network, employing depthwise separable convolutions and combining structured pruning and weight pruning strategies to simplify the network structure; S3. A knowledge distillation strategy is introduced, using a complete YOLO network pre-trained based on original road scene images as the teacher model. Through the dual constraints of feature distillation and detection distillation, image domain detection knowledge is transferred to a lightweight student model. S4. Construct a multi-objective joint loss function that integrates basic detection loss, knowledge distillation loss, and inter-class difference maximization loss. Through a progressive training strategy, achieve end-to-end collaborative optimization of the sampling matrix and the detection network, and directly complete target detection in the compressed measurement domain.

[0067] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0068] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to perform the method steps S1 to S4 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0069] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0070] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0071] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent traffic target detection method based on deep sparse representation, characterized in that: Includes the following steps: S1. Construct an adaptive compressed sensing module and process the image based on the adaptive compressed sensing module; The process includes: S11. Identify key regions of the image based on saliency detection, and dynamically allocate sampling density to the key regions and non-key regions according to a three-level hierarchical sampling strategy; S12. The image is mapped to a low-dimensional compressed representation using a learnable joint multichannel sampling matrix; S13. Optimize the feature discriminability of the compressed representation by maximizing the inter-class difference loss function, and selectively update the joint multi-channel sampling matrix based on parameter importance evaluation; S2. The obtained low-dimensional compressed representation is input into a lightweight compressed domain detection network for target recognition and localization. The lightweight compressed domain detection network adopts depthwise separable convolution and combines structured pruning and weight pruning strategies to simplify the network structure. S3. A knowledge distillation strategy is introduced for joint training. The complete YOLO network pre-trained based on the original image is used as the teacher model. Through the dual constraints of feature distillation and detection distillation, the image domain detection knowledge is transferred to the lightweight student model. S4. Construct a multi-objective joint loss function, integrating the basic detection loss, knowledge distillation loss, and inter-class difference maximization loss generated in steps S1-S3. Perform end-to-end collaborative optimization of the sampling matrix and detection network through a progressive training strategy. Finally, the optimized lightweight compressed domain detection network directly completes target detection in the compressed measurement domain and outputs the detection results.

2. The intelligent traffic target detection method based on deep sparse representation according to claim 1, characterized in that: In step S11, the saliency detection uses a method combining gradient and color contrast to calculate and obtain a saliency map; the saliency map is then normalized and Gaussian filtered before being used to guide adaptive sampling. The three-level hierarchical sampling strategy divides the image into multiple fixed-size sub-regions, performs corresponding block processing, calculates the saliency of each block through average pooling, and determines the hierarchical threshold using the statistical quantile method, thus dividing the scene image blocks into three sampling levels: high density, medium density, and low density.

3. The intelligent traffic target detection method based on deep sparse representation according to claim 2, characterized in that: In step S12, the multi-channel sampling matrix is: in, This indicates three color channels. There are three sampling density levels. Indicates the channel index. Indicates density level; Measurement Dimensions Corresponding density level: in, The total dimension of the image patch. Based on the compression sensing ratio parameter , , These are the measurement dimensions corresponding to the three sampling density levels: high, medium, and low. During the sampling process, the sub-region is separated into three color channel components: R, G, and B. Then based on the sampling level Match the sampling matrix corresponding to channel c and density identifier r. This enables the following compressed sensing sampling: in, This is the corresponding compression measurement result.

4. The intelligent traffic target detection method based on deep sparse representation according to claim 1, characterized in that: In step S13, the method for obtaining the parameter importance of the multi-channel sampling matrix is ​​as follows: ； ； ； in, For the multi-channel sampling matrix, the first The overall importance score of each parameter. , The importance of gradient and parameter magnitude are respectively. For the sampling matrix, the first Each parameter value; Total loss function For parameters The gradient represents the degree of influence of parameter changes on the loss; It is the sum of the absolute values ​​of the features in the measurement domain, reflecting the overall activation level of the current features; Based on the detection of loss, To maximize the loss for inter-class differences, , , These are feature contribution weights, task weights, and discriminative weights, respectively. Here is the numerical stability constant. These are the corresponding weighting coefficients.

5. The intelligent traffic target detection method based on deep sparse representation according to claim 1, characterized in that: In step S13, the low importance parameter is updated based on the following formula, thereby updating the joint multi-channel sampling matrix: in, For elements in the sampling matrix Current parameter value, To update the obtained parameter values, The learning rate controls the step size for updating parameters. This is the gradient of the loss function with respect to this parameter; The update mechanism of the joint multi-channel sampling matrix is ​​constructed based on key indicators such as feature variance, feature amplitude, task effectiveness, and discrimination effectiveness to evaluate the information quality of the measurement domain.

6. The target scene detection method based on deep sparse representation according to claim 1, characterized in that: In step S2, the depthwise separable convolution includes depthwise convolution and pointwise convolution, which are responsible for spatial local feature extraction and cross-channel feature fusion, respectively.

7. The intelligent traffic target detection method based on deep sparse representation according to claim 1, characterized in that: In step S2, the structured pruning strategy is based on gradient sensitivity. Weight magnitude Characteristic response intensity and task relevance Construct the following comprehensive weighted score : in, For the corresponding weighting coefficients, These are the weight parameters of the sampling matrix. The gradient vector of the loss function. Let be the partial derivative with respect to the weight parameters; Weighted pruning strategy, based on the set pruning threshold Remove weights with small values ​​to obtain sparse connections: in, For position Pruning mask, For the corresponding weight parameters; The specific pruning ratio will vary over time. The changes are shown in the following formula: in, and These represent the current training round and the total training round, respectively. The target pruning ratio.

8. The system for target scene detection based on deep sparse representation according to any one of claims 1-7, characterized in that: include: The compressed sensing module is used to build an adaptive compressed sensing module to perform saliency detection on images to identify key regions. It dynamically adjusts the sampling density based on a three-level hierarchical sampling strategy, maps high-dimensional scene images to low-dimensional compressed representations through a learnable joint multi-channel sampling matrix, enhances the discriminability of targets in the compressed domain by maximizing the inter-class difference loss function, and achieves selective updating of the sampling matrix based on parameter importance evaluation. The lightweight compressed domain detection module is used to receive the low-dimensional compressed representation, construct a lightweight network using depthwise separable convolution combined with structured pruning and weight pruning strategies, extract features in the compressed domain and output preliminary detection results. The cross-domain knowledge distillation module, based on the knowledge distillation strategy, is used to transfer the detection knowledge learned by the teacher network in the original pixel domain to the lightweight compressed domain detection module during the training phase through the dual constraints of feature distillation and detection distillation. The end-to-end collaborative optimization module is used to construct a multi-objective joint loss function that includes basic detection loss, knowledge distillation loss, and inter-class difference maximization loss. It performs collaborative optimization on the adaptive compressed sensing module and the lightweight compressed domain detection module through a progressive training strategy, and finally completes target detection directly in the compressed measurement domain.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the intelligent traffic target detection method based on deep sparse representation as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent traffic target detection method based on deep sparse representation as described in any one of claims 1-7.

Images