A target detection method and system based on a scale attention auxiliary learning method

Abstract

The application discloses a target detection method and system based on a scale attention auxiliary learning method, which enhances the detection accuracy in a complex traffic scene by explicitly emphasizing the scale-sensitive attention of feature representation, introduces an auxiliary scale estimation network, utilizes the multi-scale features of a backbone network, and is supervised through a Gaussian heat map, adopts a joint optimization strategy, simultaneously learns the auxiliary network and the object detector, thereby promoting the features sensitive to the scale change of objects, enhancing the attention of the backbone network to the multi-scale information in the network training process, improving the expression capability of the model to the scale information in the image, and removing the scale perception network in the inference stage after the training is completed, using a standard feature extraction, feature fusion and detection paradigm to complete the detection task, which improves the detection effect of the model without increasing the additional calculation burden, and there are various optimization approaches in the auxiliary network design.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving environmental perception, specifically relating to a target detection method and system based on scale attention-assisted learning. Background Technology

[0002] Accurate object detection is crucial for ensuring the safety of autonomous driving and intelligent transportation systems (ITS). In recent years, LiDAR has gradually become the main force in object detection in traffic scenes due to its accurate depth estimation capabilities. However, in some important vision tasks, such as the recognition and detection of traffic lights, traffic signs, and pedestrians, images remain indispensable because they provide rich semantic and textural information.

[0003] Traditional visual object detection methods rely heavily on hand-designed features, which often have performance limitations. Deep learning has gradually replaced these traditional methods. Deep learning-based object detection methods use convolutional neural networks (CNNs) or Transformers to extract features. In the field of intelligent transportation, CNN-based methods are widely used due to their high real-time performance and low resource consumption, while Transformer-based methods are less favored. State-of-the-art CNN-based detectors have achieved impressive accuracy and speed on general object detection datasets. However, these models often encounter difficulties when applied to challenging tasks, such as detecting multi-scale objects in complex traffic scenes. In real-world traffic scenarios, driving-related objects typically transition from far to near, creating unavoidable scale variations. Even in stationary scenes, traffic participants such as cars and pedestrians are constantly moving. Furthermore, objects of different scales can coexist in the field of view. For example, small cones and large trucks may appear in the same scene. For practical object detectors in intelligent transportation or driving systems, the "continuity" and "stability" of detection results are crucial. Therefore, addressing the challenge of large-scale changes in object detection should be an important consideration when designing visual detectors.

[0004] The challenge of addressing scale variations in complex traffic scenarios stems from the structural and learning limitations of existing general-purpose detectors. Despite employing multi-scale optimization techniques, these detectors have failed to effectively capture scale-sensitive feature representations. Research indicates that incorporating Feature Pyramid Networks (FPNs) into models can improve their sensitivity to scale information; therefore, existing detectors integrate FPNs into their models to address the multi-scale problem in detection. However, the performance improvements brought by feature pyramids primarily come from enhancing multi-layer feature maps. This approach leads to an uneven distribution of scale sensitivity among features at different scale levels and introduces excessive redundant model parameters. Furthermore, real-time performance is crucial in autonomous driving applications. Some methods increase the number of learnable parameters to enhance the detector without considering time efficiency, which may not yield optimal solutions. The successful application of auxiliary learning methods provides new insights into addressing the scale problem. These methods not only have the potential to enhance robust feature representations but also effectively maintain the real-time performance of the original framework. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a target detection method based on scale attention-assisted learning. Based on the basic idea of ​​assisted learning, an additional scale attention network is introduced during the model training stage to enhance the attention of the backbone network to multi-scale information during network training, thereby improving the model's ability to express scale information in images. In the subsequent feature fusion and prediction stages, and in the inference stage after training, the scale-aware network is removed, and the detection task is completed using standard feature extraction, feature fusion, and detection paradigms. This improves the model's detection performance without adding additional computational burden. There are multiple optimization approaches in the design of the assisted network, and the assisted network can be adjusted for specific tasks and application scenarios to cope with complex traffic scenarios.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a target detection method based on scale attention-assisted learning, which, after data augmentation of an RGB image, predicts the detection task results based on a baseline 2D target detection network, and outputs the bounding boxes of all predicted objects in the image and the categories of all objects, including the following steps:

[0007] Extracting multi-scale image features from RGB images;

[0008] The multi-scale image features are fused to obtain the fused feature representation;

[0009] Detection results are generated based on the fused feature representation;

[0010] The baseline 2D object detection network consists of a backbone network module, a feature pyramid module, and a detection head. The backbone network is used for feature extraction, the feature pyramid is used for feature fusion, and the detection head is used to predict the output results.

[0011] Furthermore, data augmentation includes cropping, flipping, and affine transformation of the input RGB image.

[0012] Furthermore, the backbone network is a CNN-based feature extractor using CSPDarkNet, which includes an attention module and four stage layers. A convolutional module with a kernel size of 3 is applied, followed by batch normalization and the Swish activation function. After the convolution, a CSPlayer is set.

[0013] Furthermore, the feature pyramid network includes multiple feature maps of different scales, forming an image pyramid network with increasing scale from the bottom to the top, where each feature map corresponds to a different resolution of the original image.

[0014] Furthermore, the detection head includes a regression branch and a classification branch. The regression branch predicts the bounding box offset from a predefined object center point p. The classification branch predicts predefined object categories and also calculates the target evaluation score factor for the detection results. The detection head utilizes encoded multi-scale fusion features, and the network within the detection head generates the final detection result. h The learnable parameters of the detection head network are {F1',F2',...,F...}. k '} represents the feature maps fused at different spatial resolutions, as shown in the following equation:

[0015]

[0016] Furthermore, real bounding boxes are used when training the detection head network. and their corresponding object category tags Supervise the model optimization. Represent a set of loss functions Used to calculate classification loss. Used to calculate regression loss. Calculate the prediction confidence loss.

[0017] Furthermore, a scale-sensitive attention network is constructed and jointly optimized by introducing it into a baseline 2D target detection network;

[0018] The optimization objective is pixel logistic regression, which includes a penalty reduction for focus loss, as shown in the following equation:

[0019]

[0020] Where α and β represent the hyperparameters of the focus loss, and N represents the image I. Λ The number of key points is determined by normalizing N to ensure that the focus loss of all positive samples is normalized to 1, based on the training settings of this branch proposed in CenterNet.

[0021] The total loss function is expressed as:

[0022]

[0023] Where λ sa λ represents the weighting coefficients of the scale-sensitive loss. cls , λ reg and λ obj The weighting coefficients representing the detection loss To predict the loss of confidence level.

[0024] Furthermore, the scale-sensitive attention network includes a feature map normalization module, a heatmap generation module, and a scale estimation module; in the feature map normalization module, the multi-scale image feature representation F... j As input, feature maps with different resolutions are downsampled and unified to a single size; the heatmap generation module uses bounding box information from annotations to generate a series of heatmaps with different scales; the scale estimation module uses convolutional layers to extract features from the normalized feature maps based on the scale prediction results of the convolutional network and generates predicted heatmaps during training.

[0025] Similar to the above-mentioned method concept, this invention provides a target detection system based on a scale attention-assisted learning method. After data augmentation of an RGB image, the system predicts the detection task results based on a baseline 2D target detection network and outputs the bounding boxes of all predicted objects on the image and the categories of all objects, including a backbone network module, a feature pyramid module, and a detection head.

[0026] The backbone network module is used to extract multi-scale image features from RGB images;

[0027] The feature pyramid module is used to fuse the multi-scale image features to obtain the fused feature representation;

[0028] The detection head generates detection results based on the fused feature representation.

[0029] Another computer device is provided, including a processor and a memory. The memory is used to store a computer executable program. The processor reads the computer executable program from the memory and executes it. When the processor executes the program, it can implement the target detection method based on the scale attention-assisted learning method of the present invention.

[0030] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the target detection method based on scale attention-assisted learning described in the present invention.

[0031] Compared with existing technologies, this invention has at least the following beneficial effects: This invention introduces an auxiliary optimization method, utilizing a separable scale estimation network to learn scale-sensitive attention, thereby enabling the feature representation in the detection network to effectively capture scale changes; it improves the model's ability to express scale information in images. In the subsequent feature fusion and prediction stages, and in the inference stage after training, the scale-aware network is removed, and the standard feature extraction, feature fusion, and detection paradigm is used to complete the detection task. This improves the model's detection performance without adding additional computational burden. Multiple optimization approaches exist in the auxiliary network design, allowing for adjustments to the auxiliary network to address complex traffic scenarios for specific tasks and application scenarios. Performance tests on multiple datasets and intelligent vehicle platforms, both qualitative and quantitative evaluation results, demonstrate that applying the method described in this invention to a state-of-the-art detector can improve detection accuracy. Real-world vehicle deployment of the detector showcases its robustness and effectiveness. Attached Figure Description

[0032] Figure 1 This paper compares the existing 2D object detection framework with the framework used in this invention.

[0033] Figure 2 An attention-assisted learning framework for target detection scale.

[0034] Figure 3 For performance comparison on the Caltech pedestrian detection dataset.

[0035] Figure 4 Deployment and testing of algorithms for intelligent vehicle platforms. Detailed Implementation

[0036] The exemplary embodiments of this application are described in detail below with reference to the accompanying drawings and specific implementations, including various details of the embodiments of this application to aid understanding. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

[0037] This invention can be flexibly integrated into any CNN-based 2D object detector. To illustrate the method provided by this invention within the overall detection framework, the YOLOX object detection model is used as an example to explain how to apply the method. First, a CNN-based baseline 2D object detector is established as the baseline detection model. Figure 2 The top-left region (during training and testing) is represented. A scale attention branch is then constructed using an auxiliary learning method, and the training process is supervised at scale using existing labels from the dataset as ground truth. After joint optimization to enhance the model's backbone network, the scale attention network is removed during inference, and the effectiveness of the application is tested on the validation set of the dataset.

[0038] Step 1: Construct a baseline 2D object detector. The baseline 2D object detector takes an RGB image as input. After data augmentation by cropping, flipping, and performing affine transformations on the input RGB image, the augmented image data is input into a convolutional neural network to predict the detection result. The baseline 2D object detector outputs the bounding boxes of all predicted objects in the image and the categories of all objects. This baseline 2D object detection network includes: a backbone network module, a feature pyramid module, and a detection head. The backbone network is used for feature extraction, the feature pyramid is used for feature fusion, and the detection head is used to predict the output result. Specific details are as follows.

[0039] Backbone network: The backbone network is represented by the symbol f. backbone It is a CNN-based feature extractor. Equation (1) illustrates the generation of deep multi-scale feature representations, F i The dimension is [C i H i W i ], applied through a convolutional layer to the input image I Λ Its dimensions are [C,H,W], and for RGB images, the number of feature channels C=3.

[0040] {F1,F2,...,F n}=f backbone (I Λ ;Θ b (1)

[0041] Where, Θ b The learnable parameter network represents the backbone network, where n equals the total number of output layers.

[0042] The commonly used backbone network CSPDarkNet employed in this invention includes a Focus module and four Stage layers. It applies a convolutional module with a kernel size of 3, followed by batch normalization and Swish activation. After convolution, CSPlayer is used to improve the CNN's learning ability and reduce memory consumption.

[0043] Feature Pyramid Network: The feature pyramid network consists of multiple feature maps, each corresponding to a different scale. Typically, the feature pyramid is an image pyramid network structure with increasing scale from the bottom to the top. Each feature map corresponds to a different resolution of the original image and contains semantic information at different levels. The feature pyramid is located between the backbone network and the detection head. The feature pyramid is used to fuse the multi-scale image features extracted by the backbone network. The feature pyramid is implemented in two ways: (1) Top-down construction: Starting from the original image, a set of feature maps is generated by downsampling multiple times. Pooling or convolution operations can be used for example downsampling. Each downsampling reduces the size of the feature map and increases semantic information. Then, the lower-level feature maps are upsampled to the same size as the higher-level feature maps by upsampling or interpolation operations to obtain a set of feature maps with different scales, which constitute the feature pyramid. (2) Bottom-up construction: Starting from the original image, a set of feature maps is generated through multiple convolution operations. Each feature map has different scales and semantic information. Subsequently, upsampling or interpolation operations are used to upsample the higher-level feature maps to the same size as the lower-level feature maps, resulting in a set of feature maps with different scales, forming a feature pyramid. The backbone network extracts image features {F1, F2, ..., F...}. n Finally, the fused feature representation {F1',F2',...,F} is obtained. k '}.

[0044] Detection Head: The detection head receives a set of feature maps fused at different spatial resolutions after passing through a pyramid network, denoted as {F1', F2', ..., F...}. k '}. In the detection head f head In the detection head, there are two key branches: a regression branch and a classification branch. By utilizing the encoded multi-scale fusion features, the network generates the final detection result, as shown in Equation (2). The regression branch predicts the bounding box offset from the predefined object center point (or object key point) p. The classification branch predicts the predefined object category Cpred_obj. Additionally, a target evaluation score factor for the detection results is calculated. Θ h This represents the learnable parameters of the detector head network.

[0045]

[0046] To facilitate parameter learning in the aforementioned detection head network, real bounding boxes are used. and their corresponding object category tags Supervise the model optimization, where Represent a set of loss functions Used to calculate classification loss. Used to calculate regression loss. Calculate the prediction confidence loss.

[0047] This invention constructs a scale-sensitive attention network for assisted learning: To improve the performance of multi-scale object detection without incurring additional inference time costs, this invention proposes a scale-sensitive attention network. An assisted learning task is constructed using two-dimensional bounding boxes from training images. The scale-sensitive attention network includes feature map normalization, heatmap generation, and scale estimation. Specific details are as follows:

[0048] Feature map normalization: Feature map normalization accepts multi-scale image feature representations F j As input, each F j The size is [C j H j W j (j = 1, 2, ..., k), where k represents the number of output levels from the FPN; using the input backbone network, by increasing the number of deconvolution modules, feature maps with different resolutions are downsampled and unified to a single size. Then, the size of the multi-level feature maps becomes... However, these are actually different representations of the input image used for detection. As an example, this invention will use C... n Set it to 128, and S to 4.

[0049] Heatmap Generation: The heatmap generation module uses bounding box information from annotations to generate a series of heatmaps with varying scales. To eliminate potential ambiguity between different scales, this invention divides the object into k scale levels based on downsampled values ​​from the backbone network. As an example, the downsampled values ​​are... Then, the k scale labels are evenly divided into l = 2k scale intervals as scale labels. The relative scale labels are calculated based on the object bounding box size of the image annotation, and a Gaussian distribution G∈[0,1] based on all objects on the data-augmented image is generated. Equation (3) illustrates the calculation of the Gaussian kernel centered on the object bounding box. The size of the Gaussian heatmap region depends on the bounding box size of the object, so that the heatmap can represent the scale of the object. If two Gaussian regions of the same scale level overlap, the maximum value is taken from the elements of the two Gaussian regions of the same scale level.

[0050]

[0051] Where (x,y) is the pixel position in the object region, and l is the level index of the scale interval, (x c ,y c ) is the center of the bounding box of each object, σ p It is the standard deviation of an object size adaptation.

[0052] Scale estimation: The scale estimation module uses a convolutional network to predict the generated scale. In scale estimation, 3×3 and 1×1 convolutional layers are used to extract features from the normalized feature map and generate a predicted heatmap during training. Within each level, the size of the predicted heatmap is... This serves as a representation for each level.

[0053] Joint optimization is performed using an auxiliary network and a standard detection network. By integrating the previously mentioned scale-sensitive attention network, the predicted image scale representation can be obtained. To ensure that the scale estimation branch affects the model but not the inference process, the auxiliary network is integrated into the overall model framework through joint optimization, as shown below.

[0054] Scale-Sensitive Attention Loss: A heatmap G is generated for each object in the image using its true label. xyl The prediction results are derived from the output of the scale estimation module. The optimization objective is pixel-wise logistic regression, which includes a penalty reduction for focus loss, as shown in Equation (4).

[0055]

[0056] Where α and β represent the hyperparameters of the focus loss, and N represents the image I. Λ The number of keypoints. By normalizing N, it can be ensured that the focus loss of all positive samples is normalized to 1. Based on the training settings of this branch proposed in CenterNet, this invention sets α to 2 and β to 4.

[0057] End-to-end learning: Scale heatmaps are generated using ground truth labels derived from target bounding boxes and image dimensions, eliminating the need for additional annotations. When the method of this invention is applied to multiple models, the loss function used in the original method is retained to calculate the classification loss. and regression loss The total loss function is expressed as formula (5).

[0058]

[0059] Where λ sa λ represents the weighting coefficients of the scale-sensitive loss. cls , λ reg and λ obj The weighting coefficients represent the detection loss. To predict the loss of confidence level.

[0060] In this invention, the coefficients of the loss function are improved by modifying the coefficient λ. cls , λ reg and λobj The values ​​were set to 1.0, 5.0, and 1.0, respectively. The impact of auxiliary learning on the detection task was explored through ablation studies, and the weighting coefficient λ of the scale-sensitive loss was examined. sa The impact of the study on attention loss at different scales is specifically manifested in: higher λ sa The values ​​[λ] show some performance improvement for detecting small and medium-sized objects. However, they negatively impact the detection of large objects, leading to a performance degradation. This performance degradation on large objects is due to the fact that increasing λ... sa When the target size is low, the model tends to prioritize handling challenging cases (detection of small and medium-sized objects). To highlight the importance of scale information in 2D images by jointly optimizing the backbone network structure, an auxiliary network is introduced. This auxiliary network estimates the scale level and contributes to the scale-sensitive loss function. The entire optimization process enables the model to learn end-to-end, thereby improving the performance of the multi-scale object detection model.

[0061] The effectiveness of the proposed method was verified on a traffic scene object detection dataset and an intelligent vehicle platform. The Caltech dataset was used for training and validation of pedestrian detection, while the Tsinghua-Tencent 100K (TT100K) dataset was used for traffic sign detection evaluation. The Caltech dataset contains 10 hours of driving video with annotations for each pedestrian in the videos. This dataset contains approximately 20,000 images with pedestrian annotations, distributed across six traffic scenes, including city streets, highways, and parking lots. TT100K is a widely used benchmark dataset for traffic sign detection tasks. TT100K contains 6,000 images for training and 3,000 images for testing. The dataset covers more than 80 categories of Chinese traffic signs. The validation platform for this invention is the "Pioneer" intelligent vehicle, which won the championship in the China Intelligent Vehicle Future Challenge in 2018, 2019, and 2020. This platform is an L4-level autonomous driving vehicle provided by the applicant.

[0062] This invention uses Faster R-CNN, SSD, RetinaNet, and AutoAssign as reference methods, and ATSS, GFL, and YOLOX as baseline object detectors. Scale attention methods are applied to these baseline detectors, resulting in SA-ATSS, SA-GFL, and SA-YOLOX, i.e., improved detectors. These models are trained using the same strategy, and the algorithm details are as follows:

[0063] Parameter definition numerical values Batchsize 8 Number of training rounds 200 Learning rate 0.001 Optimizer SGD Weight decay 0.0001 GPU NVIDIA GTX 2080 Ti Deep learning framework Pytorch Algorithm framework MMDetection 2.25.1 Deployment framework OnnxRuntime 1.7.0

[0064] The test results on the pedestrian detection and traffic sign detection datasets are shown in the table below.

[0065] Table 1. Results of the proposed method on the Caltech dataset.

[0066]

[0067] Table 2 shows the results of the proposed method on the TT100K dataset.

[0068]

[0069] In summary, this invention provides a target detection method based on scale attention-assisted learning. Based on the fundamental idea of ​​assisted learning, an additional scale attention network is introduced during the model training phase to enhance the backbone network's attention to multi-scale information, thereby improving the model's ability to express scale information in images. In the subsequent feature fusion and prediction phases, and in the inference phase after training, the scale-aware network is removed, and the detection task is completed using standard feature extraction, feature fusion, and detection paradigms. This improves the model's detection performance without adding additional computational burden. Multiple optimization approaches exist in the design of the assisted network, allowing for adjustments to the network to address complex traffic scenarios for specific tasks and applications.

[0070] To address the challenges of object detection in autonomous driving, this invention proposes a scale-aware solution. This invention constructs an auxiliary scale-aware attention network based on a general object detection framework, achieving joint optimization by combining 2D bounding box annotations from the object detection dataset with the auxiliary network. In the standard model training process, the input image data is first transformed to achieve data augmentation. The resulting image then enters a standard backbone network to extract image feature representations, followed by feature-level fusion through a pyramid network structure. Finally, the features are fed into the detection head for object category and location prediction. The auxiliary network used in this invention employs parallelism and branches between the feature pyramid and the detection head to complete scale-level supervision, improving the backbone network's ability to represent multi-scale object features. Experimental results on the Caltech pedestrian detection dataset and the TT100K traffic sign detection dataset demonstrate the effectiveness of the proposed method. Furthermore, testing the improved object detection model on an intelligent vehicle platform verifies the superiority and robustness of this invention in practical applications.

[0071] Based on the above method, the present invention provides a target detection system based on scale attention-assisted learning method. After data augmentation of RGB images, the system predicts the detection task results based on a baseline 2D target detection network and outputs the bounding boxes of all predicted objects on the image and the categories of all objects, including a backbone network module, a feature pyramid module and a detection head.

[0072] The backbone network module is used to extract multi-scale image features from RGB images;

[0073] The feature pyramid module is used to fuse the multi-scale image features to obtain the fused feature representation;

[0074] The detection head generates detection results based on the fused feature representation.

[0075] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the target detection method based on the scale attention-assisted learning method described in the present invention when executing the computer executable program.

[0076] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the target detection method based on the scale attention-assisted learning method described in the present invention.

[0077] The computer equipment may be a laptop, desktop computer, workstation, or vehicle-mounted computer.

[0078] The processor described in this invention may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf programmable gate array (FPGA).

[0079] The memory described in this invention can be an internal storage unit of a laptop, desktop computer, workstation, or vehicle-mounted computer, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0080] Computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0081] Finally, it should be noted that the above description is only for illustrating specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that any modifications or variations made based on the technical solutions and inventive concepts of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A target detection method based on scale-attention-assisted learning, characterized in that, After data augmentation of the RGB image, the detection task results are predicted based on the baseline 2D object detection network, outputting the bounding boxes of all predicted objects in the image and the categories of all objects, including the following steps: Extracting multi-scale image features from RGB images; The multi-scale image features are fused to obtain the fused feature representation; Detection results are generated based on the fused feature representation; The baseline 2D target detection network consists of a backbone network module, a feature pyramid module, and a detection head. The backbone network is used for feature extraction, the feature pyramid is used for feature fusion, and the detection head is used to predict the output results. A scale-sensitive attention network is constructed and jointly optimized by introducing it into a baseline 2D target detection network. The optimization objective is pixel logistic regression, which includes a penalty reduction for focus loss, as shown in the following equation: in, and The hyperparameter representing the focus loss, Representing an image The number of key points, through analysis Normalization is performed to ensure that the focus loss of all positive samples is normalized to 1, based on the scale-sensitive attention network training settings proposed in CenterNet. Represents a heatmap; The total loss function is expressed as: in The weighting coefficients represent the scale-sensitive loss. , and The weighting coefficients representing the detection loss To predict the loss of confidence level, Indicates regression loss, Represents classification loss, Indicates scale-sensitive loss; The scale-sensitive attention network includes a feature map normalization module, a heatmap generation module, and a scale estimation module; the feature map normalization module represents multi-scale image features. As input, feature maps with different resolutions are downsampled and unified to a single size; the heatmap generation module uses bounding box information from annotations to generate a series of heatmaps with different scales; the scale estimation module uses convolutional layers to extract features from the normalized feature maps based on the scale prediction results of the convolutional network and generates predicted heatmaps during training.

2. The target detection method based on scale attention-assisted learning according to claim 1, characterized in that, Data augmentation includes cropping, flipping, and affine transformation of the input RGB image.

3. The target detection method based on scale attention-assisted learning according to claim 1, characterized in that, The backbone network is a CNN-based feature extractor using CSPDarkNet, which includes an attention module and four Stage layers. A convolutional module with a kernel size of 3 is applied, followed by batch normalization and the Swish activation function. After the convolution, the CSPrayer is set.

4. The target detection method based on scale attention-assisted learning according to claim 1, characterized in that, The feature pyramid network consists of multiple feature maps at different scales, forming an image pyramid network with decreasing scale from the bottom to the top. Each feature map corresponds to a different resolution of the original image.

5. The target detection method based on scale attention-assisted learning according to claim 1, characterized in that, The detection head includes regression and classification branches. The regression branch starts from a predefined object center point. Predict bounding box offset The classification branch predicts predefined object categories and also calculates the target evaluation score factor for the detection results. ; The detection head utilizes encoded multi-scale fusion features, and the network within the detection head generates the final detection result. This represents the learnable parameters of the detection head network. The feature maps fused at different spatial resolutions are as follows: ； When training the detection head network, use real bounding boxes. and their corresponding object category tags Supervise the model optimization. Represent a set of loss functions, Used to calculate classification loss. Used to calculate regression loss. Calculate the prediction confidence loss.

6. A target detection system based on a scale-attention-assisted learning method, characterized in that, After data augmentation of the RGB image, the detection task results are predicted based on the baseline 2D object detection network. The output includes the bounding boxes of all objects predicted on the image and the categories of all objects, including the backbone network module, the feature pyramid module and the detection head. The backbone network module is used to extract multi-scale image features from RGB images; The feature pyramid module is used to fuse the multi-scale image features to obtain the fused feature representation; The detection head generates detection results based on the fused feature representation; A scale-sensitive attention network is constructed and jointly optimized by introducing it into a baseline 2D target detection network. The optimization objective is pixel logistic regression, which includes a penalty reduction for focus loss, as shown in the following equation: in, and The hyperparameter representing the focus loss, Representing an image The number of key points, through analysis Normalization is performed to ensure that the focus loss of all positive samples is normalized to 1, based on the scale-sensitive attention network training settings proposed in CenterNet. Represents a heatmap; The total loss function is expressed as: in The weighting coefficients represent the scale-sensitive loss. , and The weighting coefficients representing the detection loss To predict the loss of confidence level, Indicates regression loss, Represents classification loss, Indicates scale-sensitive loss; The scale-sensitive attention network includes a feature map normalization module, a heatmap generation module, and a scale estimation module; the feature map normalization module represents multi-scale image features. As input, feature maps with different resolutions are downsampled and unified to a single size; the heatmap generation module uses bounding box information from annotations to generate a series of heatmaps with different scales; the scale estimation module uses convolutional layers to extract features from the normalized feature maps based on the scale prediction results of the convolutional network and generates predicted heatmaps during training.

7. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading the computer-executable program from the memory and executing it, and the processor executing the program being able to implement the target detection method based on the scale attention-assisted learning method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the target detection method based on the scale attention-assisted learning method as described in any one of claims 1-5.

Images