Target detection model for special detection of insulation protection appliance

Abstract

The invention discloses a target detection model for special detection of an insulation protection appliance, and relates to the field of computer vision and target detection. According to the method, a multi-scale receptive field prior module and a cross-stage fusion anchor point attention module are introduced into a backbone network, so that the feature extraction capability is enhanced; meanwhile, a cross-scale interaction module is constructed in a neck network, a small target detection branch is additionally arranged at a prediction head, and multi-scale feature representation is optimized through cross-scale feature fusion. According to the method, the problems of insufficient feature extraction, small target missing detection and the like of a basic model in a complex power operation scene are solved, and the detection precision of protection appliances with different sizes is remarkably improved by enhancing the feature representation capability of a backbone network and optimizing a neck multi-scale feature fusion mechanism; and a reliable technical basis is provided for constructing an intelligent safety monitoring system.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and target detection, and in particular to a target detection model for special detection of insulating protective equipment. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] According to public information from relevant domestic and international research literature and authorized patents, existing target detection technologies often use backbone networks based on typical architectures such as CSPDarknet or ResNet, achieving feature extraction through stacked convolutional layers and cross-stage residual modules (such as CSP modules). However, due to the fixed-size convolution kernels, such designs result in a single receptive field, making it difficult to capture the multi-scale spatial distribution characteristics of insulating protective gear in complex power scenarios (e.g., the coexistence of large-sized insulating gloves at close range and small-sized goggles at long range). Meanwhile, feature fusion methods typically employ feature pyramid networks (FPNs) or their improved structures (such as PANets) to enhance semantic consistency through top-down or bottom-up unidirectional hierarchical feature fusion. However, such single-path fusion mechanisms lack deep cross-scale feature interaction, resulting in the abstraction of small target details (such as goggle edges and glove textures) in shallow high-resolution feature maps when transferred to higher layers. This makes it difficult for high-level semantic information to sink to lower-level features, ultimately leading to problems such as weakened feature expression of small targets and decreased positioning accuracy. This becomes a key bottleneck in the missed detection of protective gear in densely occluded or complex backgrounds. Summary of the Invention

[0004] The purpose of this invention is to address the need for protective gear detection in live working environments with 10kV boom trucks and propose a target detection model, INSP-YOLO (Insulation Protection YOLO), for specialized detection of insulation protective gear. This model improves YOLOv8 by optimizing the backbone network and feature fusion neck network architecture, effectively improving the detection accuracy of safety equipment such as insulating helmets, goggles, and insulating gloves. Specifically, the model introduces a Multi-Scale Receptive Field Prior Module (MS-RFPM) and a C2f-CAA module into the backbone network to enhance feature extraction capabilities. Furthermore, a Cross-Scale Interaction Block (CSIB) is constructed in the neck network, and a small target detection branch is added to the prediction head. Multi-scale feature representation is optimized through cross-scale feature fusion. This invention overcomes the problems of insufficient feature extraction and missed small target detection in complex power operation scenarios, often encountered by basic models. By enhancing the feature representation capabilities of the backbone network and optimizing the multi-scale feature fusion mechanism in the neck network, the model significantly improves the detection accuracy of protective gear of different sizes, providing a reliable technical foundation for building intelligent safety monitoring systems.

[0005] The technical solutions of the present invention are as follows:

[0006] A target detection model for specialized detection of insulating protective equipment, based on the YOLOv8 network architecture, includes:

[0007] Improved backbone network, replacing the first two C2f modules with the multi-scale receptive field prior module MS-RFPM, and replacing the last two C2f modules with the cross-stage fusion anchor attention module C2f-CAA;

[0008] The optimized neck network includes a cross-scale interaction module (CSIB) for information interaction between feature maps of different scales.

[0009] The newly added small object detection branch builds a cross-level feature interaction channel by upsampling the maximum scale feature map and concatenating it with the features output by the cross-scale interaction module CSIB.

[0010] The target detection model realizes the wearing status detection of insulating protective equipment through the above-mentioned structural improvement.

[0011] Furthermore, the multi-scale receptive field prior module MS-RFPM includes:

[0012] 1×1 point-by-point convolution kernel for information integration;

[0013] A set of parallel multi-scale deep dilated convolutions are used to capture information at different scales, and a small-kernel deep convolution is introduced at the output of the large-kernel convolution.

[0014] Horizontal and vertical depthwise strip convolutions in parallel with multi-scale depthwise dilated convolutions;

[0015] Horizontal and vertical depthwise strip convolutions are performed in parallel with multi-scale convolutions to introduce prior information about the target object.

[0016] Furthermore, the cross-stage fusion anchor attention module C2f-CAA includes:

[0017] A dual-stream feature fusion architecture is constructed based on the C2f module.

[0018] Furthermore, the dual-stream feature fusion architecture includes:

[0019] The main branch retains the gradient optimization characteristics of the original CSP structure, and the new branch uses strip-shaped separable convolution with vertical size of (2k+1×1) and horizontal size of (1×2k+1). Based on the statistical distribution of target width and height in the training set, a layer-aware convolution kernel scaling strategy is designed.

[0020] Furthermore, the convolution kernel scaling strategy includes:

[0021] In the shallow feature extraction stage, the setting of horizontal long kernel 1×9 (k=4) and expansion rate 4 is adopted, while in deep feature extraction, it is switched to the configuration of horizontal long kernel 1×5 (k=2) and expansion rate 2.

[0022] Furthermore, the cross-scale interaction module CSIB includes:

[0023] The linear mapping layer is used to flatten and splice multi-scale two-dimensional features to achieve cross-scale correlation mapping;

[0024] Multi-receptive field dilated convolution MRDC consists of multiple depth-wise separable convolutions arranged in parallel.

[0025] Furthermore, the small target detection branch includes:

[0026] Bilinear upsampling unit, responsible for bilinear upsampling of the maximum scale feature map output by the neck network;

[0027] The feature splicing layer splices and fuses the upsampled feature map with the corresponding scale feature map output by the cross-scale interaction module CSIB;

[0028] The C2f module performs multi-scale feature reorganization on the spliced ​​mixed features.

[0029] Furthermore, the prior information of the target object includes: target shape and target direction.

[0030] Furthermore, the input of the cross-scale interaction module CSIB is a set of multi-scale features extracted by the backbone network; the multi-scale two-dimensional features are flattened into one-dimensional features, and all the multi-scale one-dimensional features are concatenated to obtain the feature P c ′, P c ′Through the linear mapping layer, higher-level abstract features are extracted and the feature representation of the same target at different scales is enhanced P c ″, then P c The corresponding position in ″ is intercepted according to the scale and reshaped to obtain a new two-dimensional feature, which is finally output through multi-receptive field dilated convolution.

[0031] Furthermore, the multi-receptive field dilated convolution MRDC includes three depth-wise separable convolutions arranged in parallel.

[0032] Compared with the existing technology, the beneficial effects of the present invention are:

[0033] 1. The present invention designs a multi-receptive field prior module, which realizes global context modeling and efficient extraction of target prior features through hierarchical fusion of multi-scale deep void convolution and direction and shape-sensitive deep strip convolution (horizontal + vertical), solves the scale diversity problem of protective equipment in power scenarios, and enhances the model's feature extraction ability for multi-morphological targets while maintaining computational efficiency, providing more discriminative input representation for subsequent behavior recognition tasks.

[0034] 2. This paper designs a cross-scale interaction module, which enhances the feature aggregation capability of the neck network for small targets such as those not wearing insulating gloves and goggles through information interaction between multi-scale feature maps, thereby reducing the semantic gap between layers. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is an overall framework of target detection model for special detection of insulating protective equipment;

[0036] Figure 2 Schematic diagram of the multi-scale receptive field prior module;

[0037] Figure 3 Schematic diagram of the cross-stage fusion anchor attention module;

[0038] Figure 4 Schematic diagram of the cross-scale interaction module. DETAILED DESCRIPTION

[0039] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0040] The features and performance of the present invention are further described in detail below with reference to the embodiments.

[0041] Example 1

[0042] See also Figure 1 To better identify the protective gear worn by power workers, this embodiment proposes a target detection model (i.e., the Insulation Protection YOLO framework) for specialized detection of insulation protective gear. This model, based on the YOLOv8 network architecture, includes:

[0043] Improved backbone network, replacing the first two C2f modules with the multi-scale receptive field prior module MS-RFPM, and replacing the last two C2f modules with the cross-stage fusion anchor attention module C2f-CAA;

[0044] The optimized neck network includes a cross-scale interaction module (CSIB) for information interaction between feature maps of different scales.

[0045] The newly added small object detection branch builds a cross-level feature interaction channel by upsampling the maximum scale feature map and concatenating it with the features output by the cross-scale interaction module CSIB.

[0046] The target detection model realizes the wearing status detection of insulating protective equipment through the above-mentioned structural improvement.

[0047] First, in the backbone network, this embodiment makes two key improvements:

[0048] On the one hand, the first two C2f modules are replaced by a Multi-Scale Receptive Field Prior Module (MS-RFPM); this module uses parallel multi-scale deep dilated convolutions to capture multi-scale feature information from a global perspective; at the same time, horizontal and vertical deep strip convolutions are introduced to fuse the prior information of the target object. Compared with the traditional C2f module, MS-RFPM can better capture the contextual information in the image and extract richer semantic information by virtue of the advantages of multi-scale convolution kernels. On the other hand, in order to effectively transfer the prior information backward, the present invention designs a cross-stage fusion anchor attention module (C2f-CAA) to replace the last two C2f modules in the backbone network.

[0049] Secondly, to address the shortcomings of the neck network in feature fusion, the present invention designs a cross-scale interaction block (CSIB), which aims to achieve effective fusion and information transfer between feature maps of different scales through a cross-scale interaction mechanism, and aggregate to generate more powerful and comprehensive feature representations.

[0050] In addition, to improve the problem of poor small object detection, a small object detection branch is added. This branch helps the network more effectively capture shallow semantic information, thereby improving the detection accuracy of small objects such as those wearing goggles, not wearing goggles, and not wearing insulating gloves.

[0051] Existing target detection models are difficult to effectively deal with multi-scale targets in complex power scenarios, and their detection capabilities are obviously insufficient. The problem is that the deep feature map loses the detailed information of small targets (such as goggles and insulating gloves in protective equipment) through convolution downsampling, while the shallow features retain details but lack high-level semantic information. Therefore, this embodiment designs the MS-RFPM module (such as Figure 2 As shown in Figure 2), it extracts key feature information of targets of different scales, reduces the loss of detailed feature information, and solves problems related to multi-scale changes of targets.

[0052] In this embodiment, specifically, the multi-scale receptive field prior module MS-RFPM includes:

[0053] 1×1 point-by-point convolution kernel for information integration;

[0054] A set of parallel multi-scale deep dilated convolutions are used to capture information at different scales. A small-kernel deep convolution is introduced at the output of the large-kernel convolution to obtain a larger receptive field without increasing the amount of computation.

[0055] Horizontal and vertical depthwise strip convolutions in parallel with multi-scale depthwise dilated convolutions;

[0056] Horizontal and vertical depth strip convolutions are performed in parallel with multi-scale convolutions to introduce prior information of the target object, such as the shape and direction of the target.

[0057] In this embodiment, the entire feature extraction process of the multi-scale receptive field prior module MS-RFPM can be expressed by the following mathematical formula:

[0058] X′ M =Conv 1×1 (X M )

[0059] Among them, the input features After 1×1 convolution operation, we get

[0060]

[0061] In the above formula, X′ M After parallel multi-scale deep dilated convolution, where k (1) =5,k (2) =7,k (3) =9, then and After the depth convolution of two small kernels, we get At the same time X′ M After a horizontal 1×k l Large kernel strip depth convolution is obtained Then send it into the vertical k l ×1 strip depth convolution to obtain This strip convolution can facilitate the recognition of objects with long and thin shapes (such as goggles).

[0062]

[0063] Y M =Conv 1×1 (Z″ M +(Z″ M ⊙FH W ))

[0064] In the above formula, It will And the original features X′ retained M The output of this module is a feature map FH containing rich prior information. W and Z′ M ′ to strengthen the feature representation of the full-size target and multiply it with Z″ M The feature information with rich details is obtained by adding them together, and then the features between channels are aggregated through 1×1 point-by-point convolution to finally obtain

[0065] To address the problem of insufficient detection accuracy for long strip targets and small-scale features of safety gear (such as goggles and no goggles) in power operation scenarios, this embodiment improves the backbone network by replacing the first two C2f modules to extract target feature information and strengthen the feature representation of rectangular long strip targets. Furthermore, in order to effectively retain the feature representation of long strip targets in the extracted shallow feature map, this embodiment designs a cross-stage fusion anchor attention module C2f-CAA (see Figure 3 ). This module builds a dual-stream feature fusion architecture based on C2f: the main branch retains the gradient optimization characteristics of the original CSP structure. In order to enhance the geometric feature extraction of long strip targets, a new branch is added to dynamically screen the features of high-response areas through strip-shaped separable convolutions with vertical dimensions of (2k+1×1) and horizontal dimensions of (1×2k+1). Taking into account that the target size will change dynamically with the depth of the network layer, a hierarchical-aware convolution kernel scaling strategy is designed based on the statistical distribution of target width and height in the training set. Specifically, in the shallow feature extraction stage, a horizontal long kernel of 1×9 (k=4) with an expansion rate of 4 is used, while in deep feature extraction, it is switched to a horizontal long kernel of 1×5 (k=2) with an expansion rate of 2, so as to achieve effective capture of targets with large aspect ratios such as goggles.

[0066] The existing YOLOv8 feature fusion network has the following defects:

[0067] The simple concatenation or addition of shallow features (small object details) and deep features after upsampling (large object semantics) is difficult to achieve effective complementarity, and the traditional unidirectional feature pyramid is prone to spatial information loss. In order to make full use of the multi-scale semantics and detail feature information in the backbone network, this embodiment designs a cross-scale interaction module (CSIB). Figure 4 As shown in the figure, it aims to achieve effective fusion and information transfer between feature maps of different scales through a cross-scale interaction mechanism, and aggregate to generate more powerful and comprehensive feature representations.

[0068] In this embodiment, specifically, the cross-scale interaction module CSIB includes:

[0069] The linear mapping layer is used to flatten and splice multi-scale two-dimensional features to achieve cross-scale correlation mapping;

[0070] Multi-receptive field dilated convolution MRDC consists of multiple depth-wise separable convolutions arranged in parallel.

[0071] In this embodiment, it should be noted that the cross-scale interaction process of the cross-scale interaction module CSIB is as follows:

[0072] First, the global feature information across scales of the feature pyramid is fused through linear mapping, and the receptive field is expanded through different parallel deep convolution kernels, thereby enhancing the long-range modeling capability of the convolutional neural network. CSIB consists of a linear projection layer and a set of depthwise separable convolutions with multiple convolution kernels.

[0073] Specifically, the input of this module is a set of multi-scale features extracted by the backbone network H and W are the width and height of the model input;

[0074] Flatten the multi-scale two-dimensional features into one-dimensional features, and concatenate all the multi-scale one-dimensional features to obtain features P c ′Through the linear mapping layer, higher-level abstract features are extracted and the feature representation of the same target at different scales is enhanced P c ″, then P c The corresponding position in ″ is intercepted according to the scale and reshaped to obtain a new two-dimensional feature Finally, the output is obtained through Multi-Receptive Dilated Convolution (MRDC), and the process is as follows:

[0075] F i ′=concat(DWConv 3×3 (F i ),DWConv 5×5 (F i ),DWConv 7×7 (F i ))i=2,3,4

[0076] F″ i =Conv 1×1 (F′ i ) i=2,3,4

[0077] Feature F i First, three parallel depth-wise separable convolutions are performed to obtain rich multi-scale feature information, which is then concatenated to obtain Then use 1x1 convolution to achieve information fusion between channels

[0078] In the target detection network, the hierarchy of feature representation is significantly correlated with the depth of the network. The feature maps output by the shallow network have high-resolution characteristics and strong visual interpretability. They contain geometric prior information such as pixel-level color distribution, high-frequency texture details, and target edge contours. These shallow features are highly consistent with human visual perception. As the network level progresses, deep features are down-sampled through multiple convolutions, gradually forming high-level abstract semantic representations. The spatial details of small targets are diluted, resulting in irreversible information attenuation of key features during the pooling process. However, accurate boundary positioning and category discrimination rely on the high-frequency details and spatial structure information retained in the shallow layer. In order to enhance the perception of small targets, a large-scale detection head is introduced into the original detection framework;

[0079] In the original neck network, due to the lack of a 160×160 high-resolution feature layer, cross-level feature fusion cannot be directly achieved. Therefore, we first perform a bilinear upsampling operation on the maximum-scale feature map output by the neck network. Next, the upsampled feature map is concatenated and fused with the corresponding feature map of the same size output by the CSIB module to construct a feature interaction channel across network layers, promoting information exchange and fusion of features at different levels. Finally, the concatenated and fused hybrid features are subjected to deep multi-scale feature recombination by the C2f module to improve the model's detection performance for subtle textures, boundaries, and subtle changes, thereby more accurately localizing and recognizing small objects in complex scenes.

[0080] In this embodiment, specifically, the small target detection branch includes:

[0081] Bilinear upsampling unit, responsible for bilinear upsampling of the maximum scale feature map output by the neck network;

[0082] The feature splicing layer splices and fuses the upsampled feature map with the corresponding scale feature map output by the cross-scale interaction module CSIB;

[0083] The C2f module performs multi-scale feature reorganization on the spliced ​​mixed features.

[0084] Example 2

[0085] The training data for the target detection model for specialized detection of insulating protective equipment proposed in Example 1 primarily integrates frame-by-frame data from high-altitude work videos collected at live work sites and image sets collected online. After data cleaning, a total of 1,844 images were selected. The sample labels for each image were manually annotated and verified using Makesence annotation software, and the XML format was converted to txt format. The dataset was divided into a training set of 1,474 images and a test set of 370 images in an 8:2 ratio. The dataset includes 12 categories of sample labels, including wearing an insulating cap, not wearing an insulating cap, wearing goggles, not wearing goggles, wearing an insulating cape, not wearing an insulating cape, wearing insulating gloves, not wearing insulating gloves (including not wearing insulating gloves and wearing the wrong protective gloves), and other categories. The specific sample labels and explanations are shown in Table 1.

[0086] Table 1 Label description of the power safety equipment dataset

[0087]

[0088]

[0089] Because single-category annotation information cannot accurately determine the completeness of protective equipment worn by personnel in a scene, paired category annotations were specifically designed when constructing the dataset to ensure that the wearing of protective equipment is fully reflected. Furthermore, tools commonly used during the operation were annotated. With the help of these annotations, methods such as position distribution and area intersection and union ratio were further used to deeply analyze the positional relationship between the operating tools and personnel, as well as the interaction between different tools. Through this annotation and analysis method, we can more comprehensively understand the specific details of the scene, promptly identify potential safety hazards, and effectively improve our ability to assess the safety of the working environment.

[0090] When training the power work safety equipment dataset, the model input image size was 640×640, the batch size was set to 8, the number of training rounds was 200 epochs, and other training hyperparameters remained at the default values ​​provided by the framework. All experiments involved in this paper were conducted in the same experimental environment, which was based on the Windows 11 operating system, Python version 3.11.4, a CPU of 13th Gen Intel(R) Core(TM) i7-13620H, 16GB of memory, an RTX 4050 graphics card, the deep learning framework PyTorch 2.0.1, and CUDA Version 11.8.

[0091] Example 3

[0092] To validate the superiority of the INSP-YOLO architecture proposed in Example 1, this example conducted comparative experiments on a dataset using a selection of currently popular one-stage object detection algorithms, including YOLOv5n, YOLOv6n, YOLOv8n, YOLOv10n, YOLOv11n, and YOLOv12n. The results are shown in Table 2. As shown in the table, INSP-YOLO achieves over 2% improvement in all core metrics compared to the base model, YOLOv8n. Furthermore, by adding a small object detection layer, INSP-YOLO achieves a mAP@0.5-0.95 score of 0.63, a 2% improvement over the base model, and a mAP@0.5 score of 0.918. This improved model achieves higher detection accuracy while reducing the number of parameters, FLOPs, and model weights compared to YOLOv8n, demonstrating the effectiveness of architectural optimization in balancing accuracy and efficiency.

[0093] Table 2 Comparison of experimental results between INSP-YOLO and YOLO series algorithms

[0094]

[0095]

[0096] Table 3 compares INSP-YOLO with other mainstream object detection algorithms. The two-stage object detection algorithm, FasterR-CNN, has a relatively large number of parameters and FLOPs, achieving only a detection accuracy of 0.598 on a mAP@0.5-0.95 scale and also suffers from slow inference speed. Compared to the one-stage algorithm, SSD512, INSP-YOLO achieves improvements of 13% in F1 score and 25.8% in mAP@0.5-0.95. Compared to the anchor-free detection algorithm FCOS, our proposed algorithm achieves higher accuracy with lower parameters and complexity. RT-DETR-R50, based on the Transformer architecture, achieved the highest accuracy on the dataset. However, INSP-YOLO-P2, while reducing its parameters and FLOPs by 81% and 89%, achieved improvements of 3.5% in recall and 5% in mAP@0.5-0.95, respectively. Experimental results show that INSP-YOLO achieves higher protective equipment detection accuracy in power scenarios while maintaining low parameter count and complexity.

[0097] Table 3 Comparison of INSP-YOLO and mainstream target detection algorithms

[0098]

[0099] To evaluate the effectiveness of the proposed improved method, several small-sized object categories with poor F1 score and mAP@0.5 performance, including goggle, no-goggle, and no-gloves, were selected for comparative analysis. The experimental results are shown in Table 4. INSP-YOLO demonstrates advantages in small object detection tasks: Compared with the baseline model YOLOv8n, INSP-YOLO improves both F1 score and mAP@0.5 by 7.5% in goggle detection. By adding a small object detection layer, INSP-YOLO-P2 improves F1 score and mAP to 0.786 and 0.855 respectively. In no-goggle detection, INSP-YOLO-P2's mAP@0.5 improves by 11.5% compared to YOLOv8n, and in no-gloves detection, INSP-YOLO-P2's mAP@0.5 improves by 9.5% compared to YOLOv8n. Compared to other YOLO models, YOLOv10n's lightweight design leads to insufficient feature extraction capabilities. Its mAP in no-goggle detection is only 0.607, significantly lower than INSP-YOLO's 0.71, highlighting its limitations in small-target scenarios in power grid scenarios. INSP-YOLO outperformed the baseline model in all metrics except for a slight decrease in the F1 score for the no-gloves category in three categories of small objects, validating the effectiveness of the improved model in small-target detection.

[0100] Table 4 Comparison of INSP-YOLO and mainstream target detection algorithms

[0101]

[0102] To analyze the effectiveness of each module, the contribution of the improved structure to the overall model performance is evaluated in Table 5. The results show that the activation of the MS-RFPM module effectively improves the F1 Score and mAP@0.5, reduces the number of parameters, and increases inference speed. The introduction of this module improves the backbone feature extraction capability. When the C2f-CAA attention module is further introduced, the F1 Score is further optimized to 0.876, and the mAP@0.5 is slightly improved, enhancing the feature expression capability. After further superimposing the CSIB module, the F1 Score and mAP@0.5 are further improved to 0.884 and 0.906, respectively, proving the effectiveness of the cross-stage interaction mechanism. Finally, after the P2 branch is activated, by increasing the number of parameters and sacrificing some detection speed, the F1 Score and mAP@0.5 increase to 0.889 and 0.918, respectively. At the same time, the FPS of 98 still meets the requirements of real-time detection, and effectively reduces missed detections and false detections.

[0103] Table 5 INSP-YOLO ablation experiment

[0104] MS-RFPM C2f-CAA CSIB P2 branch F1 Score mAP@0.5 Params / MiB FPS × × × × 0.857 0.883 3.2 156 √ × × × 0.870 0.895 2.98 184 √ √ × × 0.876 0.9 3.2 165 √ √ √ × 0.884 0.905 3.3 162 √ √ √ √ 0.889 0.918 4.4 98

[0105] To more comprehensively demonstrate the differentiated improvement effects of each module in INSP-YOLO on small safety equipment detection (see Table 6 for details), the basic model YOLOv8n exhibits uneven performance in small object detection: although it performs best in the "no goggles" category, it misses detections in categories such as "goggles" and "not wearing insulating gloves" due to factors such as reflective materials and small objects. First, by adding MS-RFPM, we effectively alleviated the difficulty in identifying goggles, improving their mAP@0.5 by 2.3%. Furthermore, by using C2f-CAA to enhance the model's feature selection ability in complex backgrounds, the no_goggle detection accuracy increased by 2.9%. Then, by adding CSIB through cross-layer feature fusion, the feature expression ability of goggles was further improved, increasing the mAP@0.5 of goggle to 0.803. Finally, by adding a P2-branch shallow detection head, the high-resolution feature map was used to enhance the small target response capability, significantly improving the detection accuracy and recall rate of three types of small-sized targets: goggles, gloves, and no goggles, effectively overcoming the problem of missed small target detection.

[0106] Table 6 Impact of INSP-YOLO modules on small-size targets

[0107]

[0108] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

[0109] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of the description in this section that did not constitute prior art at the time of filing are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A target detection model for special detection of insulating protective equipment, characterized in that: Based on the YOLOv8 network architecture, including: Improved backbone network, replacing the first two C2f modules with the multi-scale receptive field prior module MS-RFPM, and replacing the last two C2f modules with the cross-stage fusion anchor attention module C2f-CAA; The optimized neck network includes a cross-scale interaction module (CSIB) for information interaction between feature maps of different scales. The newly added small object detection branch builds a cross-level feature interaction channel by upsampling the maximum scale feature map and concatenating it with the features output by the cross-scale interaction module CSIB. The target detection model realizes the wearing status detection of insulating protective equipment through the above-mentioned structural improvement.

2. A target detection model for special detection of insulating protective equipment according to claim 1, characterized in that: The multi-scale receptive field prior module MS-RFPM includes: 1×1 point-by-point convolution kernel for information integration; A set of parallel multi-scale deep dilated convolutions are used to capture information at different scales, and a small-kernel deep convolution is introduced at the output of the large-kernel convolution. Horizontal and vertical depthwise strip convolutions in parallel with multi-scale depthwise dilated convolutions; Horizontal and vertical depthwise strip convolutions are performed in parallel with multi-scale convolutions to introduce prior information about the target object.

3. The target detection model for special detection of insulating protective equipment according to claim 1, characterized in that: The cross-stage fusion anchor attention module C2f-CAA includes: A dual-stream feature fusion architecture is constructed based on the C2f module.

4. A target detection model for special detection of insulating protective equipment according to claim 3, characterized in that: The dual-stream feature fusion architecture includes: The main branch retains the gradient optimization characteristics of the original CSP structure, and the new branch uses strip-shaped separable convolution with vertical size of (2k+1×1) and horizontal size of (1×2k+1). Based on the statistical distribution of target width and height in the training set, a layer-aware convolution kernel scaling strategy is designed.

5. The target detection model for special detection of insulating protective equipment according to claim 4, characterized in that: The convolution kernel scaling strategy includes: In the shallow feature extraction stage, the setting of horizontal long kernel 1×9 (k=4) and expansion rate 4 is adopted, while in the deep feature extraction stage, it is switched to the configuration of horizontal long kernel 1×5 (k=2) and expansion rate 2.

6. The target detection model for special detection of insulating protective equipment according to claim 1, characterized in that: The cross-scale interaction module CSIB includes: The linear mapping layer is used to flatten and splice multi-scale two-dimensional features to achieve cross-scale correlation mapping; Multi-receptive field dilated convolution MRDC consists of multiple depth-wise separable convolutions arranged in parallel.

7. The target detection model for special detection of insulating protective equipment according to claim 1, characterized in that: The small target detection branch includes: Bilinear upsampling unit, responsible for performing bilinear upsampling on the maximum scale feature map output by the neck network; The feature splicing layer splices and fuses the upsampled feature map with the corresponding scale feature map output by the cross-scale interaction module CSIB; The C2f module performs multi-scale feature reorganization on the spliced ​​mixed features.

8. The target detection model for special detection of insulating protective equipment according to claim 2, characterized in that: The prior information of the target object includes: the shape and direction of the target.

9. The target detection model for special detection of insulating protective equipment according to claim 6, characterized in that: The input of the cross-scale interaction module CSIB is a set of multi-scale features extracted by the backbone network; the multi-scale two-dimensional features are flattened into one-dimensional features, and all the multi-scale one-dimensional features are spliced ​​to obtain the feature P c ′, P c ′Through the linear mapping layer, higher-level abstract features are extracted and the feature representation of the same target at different scales is enhanced P c ″, then P c The corresponding position in ″ is intercepted according to the scale and reshaped to obtain a new two-dimensional feature, which is finally output through multi-receptive field dilated convolution.

10. The target detection model for special detection of insulating protective equipment according to claim 6, characterized in that: The multi-receptive field dilated convolution MRDC includes three depth-wise separable convolutions arranged in parallel.

Images