A target detection method based on a lightweight multi-scale detection perception enhancement module

Abstract

This invention discloses a target detection method based on a lightweight multi-scale detection and perception enhancement module, belonging to the field of computer vision technology. The method includes: preparing and preprocessing relevant datasets; constructing a lightweight multi-scale detection and perception enhancement module, including a receptive field enhancement module (FRFM), a position-intelligent enhanced convolutional module (PSConv), and a multi-scale attention enhancement module (MLA); applying this module to a YOLOv10 network to form a target detection model; training and validating the target detection model using the dataset; and detecting the target based on the trained target detection model. This invention introduces a lightweight multi-scale detection and perception enhancement module that can be plugged and played into various convolutional networks, enabling the detection performance to perfectly replace the original network architecture. It allows target detection tasks to be completed on low-computing devices, achieving better experimental results at a lower cost, thus contributing to the advancement of target detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a target detection method based on a lightweight multi-scale detection and perception enhancement module. Background Technology

[0002] With the development of deep convolutional neural networks, related algorithms have made significant progress in tasks such as object detection, image classification, and language processing, and have been widely applied. Therefore, higher demands are being placed on algorithm performance. Taking object detection as an example, its evaluation criteria are no longer limited to simple recognition, but have expanded to multiple dimensions such as detection accuracy, network structure design, feature representation ability, receptive field range, and the integration of attention mechanisms. These improvements aim to address the challenges in practical applications, including improving detection accuracy, enhancing the ability to recognize low-resolution targets, and improving robustness to occluded targets.

[0003] Although Transformers and their variants have been explored in various fields, many researchers still prefer Convolutional Neural Networks (CNNs) due to their large number of parameters and high computational complexity. One of the advantages of CNNs is their ability to efficiently learn from datasets of varying sizes through supervised and semi-supervised learning, adapting to diverse task requirements. However, achieving optimal performance on computationally limited devices remains a significant challenge for researchers. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention employs a convolutional network based on a lightweight multi-scale detection and perception enhancement module to solve related problems in existing multi-scale target recognition and detection tasks.

[0005] To achieve the above objectives, this invention provides a target detection method based on a lightweight multi-scale detection and perception enhancement module, comprising the following steps: (1) Prepare the relevant dataset and perform preprocessing; (2) Construct a lightweight multi-scale detection perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhancement convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model; The target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of its backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling part of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module. (3) Train and validate the target detection model using the dataset; (4) Based on the trained target detection model, the target to be detected is detected.

[0006] Furthermore, the preprocessing involves preprocessing the dataset using data enhancement methods such as mosaic processing, flipping, and translation, thereby enabling the detected images to present richer feature information in terms of content and layout.

[0007] Furthermore, the lightweight multi-scale detection and perception enhancement module can also be applied to KITTI, YOLOv11, or YOLOv8 networks to form target detection models.

[0008] Furthermore, the FRFM specifically refers to: The input feature map is sequentially processed through 1*1 convolution and dilated convolution to extract the feature information of the target and generate the convolutional feature map F. The convolutional feature map F is combined with an improved focus-aware attention mechanism, and then fused with residual features to obtain enhanced target features for output. in: This indicates improved focus-aware attention; Cat represents the fusion module; and For max pooling and average pooling; For activation functions; This indicates convolution, with the subscript indicating the size of the convolution kernel. Indicates the input feature map; Represents a multilayer perceptron; express.

[0009] Furthermore, PSConv specifically refers to: The input feature map is sequentially processed through global average pooling and a multilayer perceptron to output dynamic feature weights. The meshes in the X and Y directions are generated as follows: By combining the feature weights with the networks in the X and Y directions, feature generation in the X and Y directions is obtained; The generated features are then fused and output. in: Indicates the input feature map; Indicates global average pooling; Represents a multilayer perceptron; Indicates the width of the input feature map; Indicates the length of the input feature map; This represents a function used to generate a linearly bisected vector; Indicates the fusion module; Represents convolution; This indicates the output of PSConv; This represents the first vector output by the multilayer perceptron. This represents the second vector output by the multilayer perceptron, corresponding to the X and Y directions.

[0010] Furthermore, the MLA specifically refers to: The target features are traversed using one-dimensional convolutions on the feature maps in the width and height directions of the input feature map. Iterate through all features and output them to the filter to enhance feature details in any direction; in, It is the output after loop fusion. This is the result of multiple one-dimensional convolution traversals. i From 1... n .

[0011] Further validation metrics include: mAP50%, mAP50-95%, number of parameters, computational cost, and FPS.

[0012] The present invention also provides a target detection device based on a lightweight multi-scale detection and perception enhancement module, for performing the target detection method based on the lightweight multi-scale detection and perception enhancement module as described above, comprising: The data acquisition module is used to prepare relevant datasets and perform preprocessing. The model building module is used to construct a lightweight multi-scale detection perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhanced convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model. The target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of its backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling part of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module. The training module is used to train the object detection model using the dataset; The detection module is used to detect the target to be detected based on the trained target detection model.

[0013] The beneficial effects of this invention are: This invention introduces a lightweight multi-scale detection-based perceptual enhancement module that can be plug-and-play applied to various convolutional networks. FRFM is a novel module combining Focused Aware Attention (FAA) and parallel dilated convolution, overcoming the limitations of traditional parallel structures in feature extraction. This method expands the network's effective receptive field while accurately focusing on key regions. The PSConv module integrates global average pooling to capture contextual dependencies, and a grid combined with MLP and dynamic weight generation mechanisms adaptively enhances the target region response while actively suppressing background interference, thereby improving the spatial selectivity and discriminability of features. The MLA module achieves dynamic weighted fusion of high-level semantics and low-level details, enhancing the network's robustness to scale changes. These three modules work together in different ways, enabling the detection performance to perfectly replace the original network architecture, completing object detection tasks on low-computing devices. The main purpose is to achieve better experimental results at a lower cost, contributing to the advancement of object detection. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the target detection method based on a lightweight multi-scale detection and perception enhancement module according to an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the target detection model structure in an embodiment of the present invention (based on VOLO v10).

[0016] Figure 3This is a schematic diagram of the existing VOLO v10 network structure.

[0017] Figure 4 This is a schematic diagram of the FRFM structure according to an embodiment of the present invention.

[0018] Figure 5 This is a thermal diagram of the output of each layer in the FAA structure of the FRFM in this embodiment of the invention.

[0019] Figure 6 This is a comparison of the effective receptive field maps of FRFM with other feature fusion layers in an embodiment of the present invention.

[0020] Figure 7 This is a heatmap comparison of PSConv with other convolutions in the feature enhancement stage according to an embodiment of the present invention.

[0021] Figure 8 This is a heatmap comparison of MLA with other attention modules in an embodiment of the present invention.

[0022] Figure 9 The figure shows the generalization experiment results of the present invention on the dataset.

[0023] Figure 10 The figure shows the generalization experiment results of the embodiments of the present invention with different model structures. Detailed Implementation

[0024] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments.

[0025] like Figure 1 As shown, this embodiment of the invention provides a target detection method based on a lightweight multi-scale detection and perception enhancement module, characterized by the following steps: S101. Prepare the relevant dataset and perform preprocessing.

[0026] The datasets used in this embodiment of the invention are the VisDrone and KITTI datasets. Data augmentation strategies such as mosaic processing, flipping, and translation are used to preprocess the original datasets, thereby enabling the images to present richer feature information in terms of content and layout.

[0027] S102. Construct a lightweight multi-scale detection and perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhanced convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model. The target detection model can be based on the YOLOv10 network, or it can be based on the KITTI, YOLOv11 or YOLOv8 network, and can be enhanced by introducing a lightweight multi-scale detection perception enhancement module.

[0028] Taking the YOLOv10 network as an example, the object detection model is as follows: Figure 2 As shown, the original YOLOv10 network is as follows: Figure 3 As shown, the target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of the backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module.

[0029] like Figure 4 As shown, FRFM specifically refers to: The input feature map is sequentially processed through a combination of 1*1 convolution and dilated convolution to extract the feature information of the target and generate the convolutional feature map F.

[0030] The convolutional feature map F is combined with an improved focus-aware attention mechanism, and then fused with residual features to obtain enhanced target features for output. in: This indicates improved focus-aware attention; Cat represents the fusion module; and For max pooling and average pooling; For activation functions; This indicates convolution, with the subscript indicating the size of the convolution kernel. Indicates the input feature map; Represents a multilayer perceptron; express.

[0031] The heat map of each layer in the FAA structure is as follows: Figure 5 As shown. The effective receptive field map of FRFM compared to other feature fusion layers, for example... Figure 6 As shown.

[0032] PSConv specifically refers to: The input feature map is sequentially processed through global average pooling and a multilayer perceptron to output dynamic feature weights.

[0033] The meshes in the X and Y directions are generated as follows: By combining the feature weights with the networks in the X and Y directions, feature generation in the X and Y directions is obtained; The generated features are then fused and output. in: Indicates the input feature map; Indicates global average pooling; Represents a multilayer perceptron; Indicates the width of the input feature map; Indicates the length of the input feature map; This represents a function used to generate a linearly bisected vector; Indicates the fusion module; Represents convolution; This indicates the output of PSConv; This represents the first vector output by the multilayer perceptron. This represents the second vector output by the multilayer perceptron, corresponding to the X and Y directions.

[0034] A heatmap comparison of PSConv with other convolutions during the feature enhancement stage, for example... Figure 7 As shown.

[0035] MLA specifically refers to: The target features are traversed using one-dimensional convolutions on the feature maps in the width and height directions of the input feature map.

[0036] Iterate through all features and output them to the filter to enhance feature details in any direction.

[0037] in, It is the output after loop fusion. This is the result of multiple one-dimensional convolution traversals. i From 1... n .

[0038] Heatmap comparison of MLA with other attention modules, for example Figure 8 As shown.

[0039] S103. Train and validate the object detection model using the dataset.

[0040] Validation metrics include: mAP50%, mAP50-95%, number of parameters, computational cost, and FPS.

[0041] S104. Based on the trained target detection model, detect the target to be detected.

[0042] This invention also provides a target detection device based on a lightweight multi-scale detection and perception enhancement module, used to execute the target detection method based on the lightweight multi-scale detection and perception enhancement module described above, including: The data acquisition module is used to prepare relevant datasets and perform preprocessing. The model building module is used to construct a lightweight multi-scale detection perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhanced convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model. The target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of its backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling part of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module. The training module is used to train the object detection model using the dataset; The detection module is used to detect the target to be detected based on the trained target detection model.

[0043] The verification process of this invention embodiment is given below. The model training process is run under the Python, Linux 6.8.0, and PyTorch frameworks. The software environment is CUDA 12.8 and Python 3.12.7. The GPU used for training is an NVIDIA GeForce RTX 2080. The initial input image size is 640×640×3. The datasets selected in the scheme are VisDrone and KITTI datasets. The network frameworks selected for the experiment are YOLOv8, YOLOv10, and YOLOv11.

[0044] First, an ablation experiment was conducted. Based on YOLOv10, the experimental results of adding new modules are shown in Table 1, where ours represents the results of the proposed solution.

[0045] Table 1 The experimental results show that the overall improvement effect is significantly improved, with an increase of 1.4% in target detection accuracy, 1.3% in mAP50%, and 0.9% in mAP50-95%.

[0046] As can be seen, the new network design has a very good detection and recognition effect in target detection. It can completely replace the original model for target detection on low computing power devices. Not only has the accuracy been improved, but the recall rate also shows that the ability to avoid missed detections and false detections has been significantly improved.

[0047] Table 1 shows the experimental results for experimental groups 2, 3, 4 and 8. After 100 rounds of experiments, each module enriched the information of target features and increased the probability of correct detection in terms of receptive field, target localization and multi-scale target detection. It also has the ability to run and deploy on low computing power devices such as RTX2080.

[0048] Experimental results demonstrate that the proposed design module significantly improves accuracy on small object detection datasets and substantially enhances regression capabilities for missed and false detections. It effectively replaces the original network in various convolutional network depths, offering superior feature extraction and localization performance.

[0049] Comparative experiments between convolutional networks of different depths were used to verify the plug-and-play performance of the new module, with 100 epochs. Figure 10 As shown in Table 2 (News(3) indicates the introduction of 3 new modules).

[0050] Table 2 As can be seen from the values ​​in the table, the designed module significantly improves the mean accuracy (mAP) and precision (P) under different network structure conditions, exhibits better feature prediction performance, and is more friendly to small target recognition. This proves that the designed module can replace the original network to perform accurate target detection on devices with limited computing power, including scenarios such as occlusion, false detection, and small targets.

[0051] Performance validation results on other datasets are as follows: Figure 9 As shown in Table 3, the number of Epochs is 200.

[0052] Table 3 The experiment showed that the mAP value was improved by 2.9% and the accuracy was improved by 2.2% on the KITTI dataset, indicating that the three plug-and-play modules have strong generalization ability in handling targets in complex traffic and drone scenarios.

[0053] Based on the above examples, the algorithm module proposed in this invention improves accuracy, accurately detects occluded targets, and is suitable for different network structures. This invention has significant advantages in detecting small objects, especially in multi-layered perceptual fusion, convolution, and feature extraction. It also has important practical value for meeting the requirements of lightweight networks, low computational power, and small object detection.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target detection method based on a lightweight multi-scale detection and perception enhancement module, characterized in that, Includes the following steps: (1) Prepare the relevant dataset and perform preprocessing; (2) Construct a lightweight multi-scale detection perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhancement convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model; The target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of its backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling part of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module. (3) Train and validate the target detection model using the dataset; (4) Based on the trained target detection model, the target to be detected is detected.

2. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that, The preprocessing involves using data enhancement methods such as mosaic processing, flipping, and translation to preprocess the dataset, thereby enabling the detected images to present richer feature information in terms of content and layout.

3. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that: The lightweight multi-scale detection and perception enhancement module can also be applied to KITTI, YOLOv11, or YOLOv8 networks to form target detection models.

4. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that, The FRFM specifically refers to: The input feature map is sequentially processed through 1*1 convolution and dilated convolution to extract the feature information of the target and generate the convolutional feature map F. The convolutional feature map F is combined with an improved focus-aware attention mechanism, and then fused with residual features to obtain enhanced target features for output. in: This indicates improved focus-aware attention; Cat represents the fusion module; and For max pooling and average pooling; For activation functions; This indicates convolution, with the subscript indicating the size of the convolution kernel. Indicates the input feature map; Represents a multilayer perceptron; express.

5. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that, Specifically, PSConv refers to: The input feature map is sequentially processed through global average pooling and a multilayer perceptron to output dynamic feature weights. The meshes in the X and Y directions are generated as follows: By combining the feature weights with the networks in the X and Y directions, feature generation in the X and Y directions is obtained; The generated features are then fused and output. in: Indicates the input feature map; Indicates global average pooling; Represents a multilayer perceptron; Indicates the width of the input feature map; Indicates the length of the input feature map; This represents a function used to generate a linearly bisected vector; Indicates the fusion module; Represents convolution; This indicates the output of PSConv; This represents the first vector output by the multilayer perceptron. This represents the second vector output by the multilayer perceptron, corresponding to the X and Y directions.

6. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that, The MLA specifically refers to: The target features are traversed using one-dimensional convolutions on the feature maps in the width and height directions of the input feature map. Iterate through all features and output them to the filter to enhance feature details in any direction; in, It is the output after loop fusion. This is the result of multiple one-dimensional convolution traversals. i From 1... n .

7. The target detection method based on a lightweight multi-scale detection and perception enhancement module according to claim 1, characterized in that, Validation metrics include: mAP50%, mAP50-95%, number of parameters, computational cost, and FPS.

8. A target detection device based on a lightweight multi-scale detection and perception enhancement module, used to execute the target detection method based on a lightweight multi-scale detection and perception enhancement module as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to prepare relevant datasets and perform preprocessing. The model building module is used to construct a lightweight multi-scale detection perception enhancement module, including a receptive field enhancement module FRFM, a location-intelligent enhanced convolution PSConv, and a multi-scale attention enhancement module MLA; and apply it to the YOLOv10 network to form an object detection model. The target detection model is based on the YOLOv10 network. The C2F and C2FCIB modules of its backbone network and the C2F and C2FCIB modules of the downsampling part of the neck network are all replaced with the receptive field enhancement module FRFM. The output from the backbone network to the neck network and the output from the upsampling to the downsampling part of the neck network need to be connected to the subsequent modules after passing through the position intelligent enhancement convolution PSConv. A multi-scale attention enhancement module MLA is introduced before the first upsampling module of the neck network, after all C2F modules and after the first FRFM module. The training module is used to train the object detection model using the dataset; The detection module is used to detect the target to be detected based on the trained target detection model.

Images