Recognition method for lightweight underground irregular-shaped protective equipment

Abstract

The invention relates to the technical field of image recognition, in particular to a lightweight underground irregular shape protection equipment recognition method, which comprises the steps of data acquisition; a target recognition model is constructed to recognize target features in image data in the data set, the target recognition model is based on a YOLOv11 model, a backbone network of the target recognition model is a star-shaped backbone network, a neck network designs a double-branch CAA and a DSConv module, the double-branch CAA establishes global semantic association through upper and lower anchor point attention, and the DSConv module is adaptive to a target shape through snakelike dynamic convolution; and training and verifying the improved model by using a training set and a verification set, and identifying the underground miner personal protection equipment. According to the scheme, the backbone network and the neck are improved based on the YOLOv11 model in combination with the StarNet, the CAA and the DSConv, the lightweight PPE detection model based on the improved YOLOv11 is established, it can be guaranteed that the detection precision is improved under the complex and changeable coal mine environment, network lightweight is achieved, and the operation speed and efficiency of the model are improved.

Description

Technical Field

[0001] The present application relates to the field of image recognition technology, and in particular to a method for identifying lightweight underground protective equipment with irregular shapes. Background Art

[0002] In the complex and ever-changing mining environment, safety accidents frequently occur, posing a serious threat to miners' lives. To improve miner safety, video surveillance is often used to monitor the wearing of personal protective equipment (PPE) by underground miners during work. However, due to the complex underground mine environment, high dust concentrations, and poor visibility, existing image recognition models face difficulties in underground feature extraction, large parameter counts, poor real-time performance, and the risk of missed or false detection of small objects, resulting in low overall detection accuracy. Summary of the Invention

[0003] The technical problem to be solved by the present invention is that the existing detection model has a low detection accuracy in extreme underground environments.

[0004] To this end, the present invention provides a method for identifying lightweight underground irregular-shaped protective equipment.

[0005] The technical solution adopted by the present invention to solve its technical problem is: A method for identifying lightweight underground irregular-shaped protective equipment includes: Step 1: Data collection: divide the collected data into different categories and divide the data set into training set, validation set, and test set; Step 2: Build a target recognition model to identify target features in the image data in the dataset. The target recognition model is based on the YOLOv11 model. Its backbone network is a star-shaped backbone network. The neck network is designed with dual-branch CAA and DSConv modules. The dual-branch CAA establishes global semantic associations through upper and lower anchor point attention. The DSConv module uses dynamic kernel snake-shaped dynamic convolution to adapt to the target shape. Step 3: Use the training set and validation set to train and validate the improved model; Step 4: Use the trained model to identify the personal protective equipment of underground miners.

[0006] Furthermore, in the step 2, the star-shaped backbone network is divided into four main stages. In the first stage, large-kernel depthwise separable convolution is used to capture long-distance spatial dependencies, compensate for the local limitations of small-kernel convolution, and fuse multi-scale features using element-by-element multiplication operations; stages 2 and 3 further downsample to perform multi-scale feature extraction and spatial optimization; stage 4 captures the global context for the last time through DWConv to ensure the integrity of high-level features, and outputs feature maps in combination with global average pooling or fully connected layers.

[0007] Furthermore, in step 2, the dual-branch CAA includes parallel global and local branches, where the high-resolution local branch focuses on capturing fine details of miners and their equipment, and the low-resolution global branch reduces the feature map resolution through downsampling operations to extract and understand global spatial context information.

[0008] Furthermore, in the step 2, the global branch and the local branch perform feature enhancement through local detail attention and global layout attention respectively and then fuse the features.

[0009] Furthermore, the local branch generates low-resolution features , then use 1×1 convolution to compress the channel dimension to C / 2 to reduce the amount of calculation. Next, use DWConv with a kernel size of 1×(11+2n) to capture long-distance dependencies in the horizontal direction and output features , apply DWConv with the same kernel size again, process the vertical context, and output features Finally, the horizontal and vertical features are spliced ​​or superimposed, and the channel dimension is restored to C through the second 1×1 convolution. The spatial attention weight is generated by the Sigmoid function to weight the original feature map, and the enhanced feature map is finally output. , where l represents the current layer and l-1 represents the output of the previous layer as the input of the current layer.

[0010] Furthermore, in the step 2, the structure of the DSConv of the neck network includes an input feature map stage, a learning deformation stage, a snake convolution stage, an adaptive structural feature deformation stage and an output stage.

[0011] Furthermore, in the step 2, in the neck network, a conditional normalization layer is designed based on DSConv to modulate the generation parameters of the dynamic kernel according to the real-time environment characteristics, and the dynamic kernel is applied to the snake convolution.

[0012] Furthermore, the learning deformation stage learns deformation parameters according to the characteristics of the input feature map, provides personalized deformation guidance for the subsequent snake convolution operation, applies the learned deformation to the input feature map, and performs a snake convolution operation. After the snake convolution operation, an intermediate feature map is obtained, which contains the preliminary extracted feature information, and then adaptive structural feature deformation is performed to further optimize the feature expression, and finally the output feature map is obtained.

[0013] Furthermore, in the conditional normalization layer, first, the downhole light intensity value E is obtained in real time through the light sensor. light , use the dust sensor to measure the dust concentration in real time, and get the dust concentration estimate Edust ,Secondly, a fully connected conditional network is designed to take the real-time environment features E={E light ,E dust} as input and output the conditional information used to modulate the dynamic kernel generation parameters. After introducing the conditional normalization layer, the scaling factor γ∈R is generated by the conditional vector C. m and the offset factor β∈R m , perform affine transformation on the dynamic kernel generation parameters to obtain the dynamic kernel generation parameters θ after affine transformation cond , then, the modulated dynamic kernel is used to generate the parameters θ cond , according to the input feature map, a dynamic convolution kernel K that adapts to the current environmental conditions is generated. Finally, the generated dynamic convolution kernel K is used to perform a convolution operation on the input feature map X to obtain the output feature map O.

[0014] The beneficial effects of the present invention are as follows: based on the YOLOv11 model, this solution combines StarNet, CAA, and DSConv to improve the backbone network and neck, respectively, thereby establishing a lightweight PPE detection model based on the improved YOLOv11. StarNet is used to improve the backbone network, and its attention mechanism is used to process feature maps, enabling the model to automatically learn more discriminative features and remove unimportant feature information, thereby improving the model's feature extraction capabilities. CAA and DSConv are used to optimize the neck. A "local-global" dual-branch CAA is designed and integrated into the neck network. The attention mechanism adaptively allocates weights for different receptive fields, enabling the model to simultaneously focus on local details and global information, enhancing feature representation. DSConv uses dynamic paths to adapt to target geometry, enhancing the modeling capability of complex shapes while suppressing background noise interference. The design of the conditional normalization layer enables the generation of dynamic convolution kernels to better adapt to environmental conditions such as underground lighting and dust fluctuations, thereby improving the performance and stability of the model in practical applications. The above technologies can ensure the improvement of detection accuracy and lightweight network in the complex and changeable coal mine environment, thereby improving the running speed and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings and examples.

[0016] Figure 1 It is a flow chart of data set acquisition and processing in the present invention.

[0017] Figure 2 It is a structural diagram of the target recognition model constructed based on the YOLOv11 model in the present invention.

[0018] Figure 3 It is a structural diagram of the star backbone network in the present invention.

[0019] Figure 4 Schematic diagram of the double-branch structure in the neck network of the present invention.

[0020] Figure 5 This is the CAA context anchor attention structure diagram in the present invention.

[0021] Figure 6 It is a schematic diagram of the structure of DSConv in the present invention.

[0022] Figure 7 Schematic diagram of the change of DSConv receptive field in the present invention.

[0023] Figure 8 This is a comparison chart of the recognition results before and after the model improvement in the present invention. DETAILED DESCRIPTION

[0024] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0025] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, features defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0026] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0027] A lightweight method for identifying irregularly shaped underground protective equipment is proposed. To address background noise interference and missed small targets caused by the dense underground equipment, a star network architecture (StarNet) is introduced to replace the original YOLOv11 backbone network. This reduces computational complexity while maintaining detection accuracy, improving the algorithm's real-time performance. Secondly, we optimize for feature blurring caused by dust scattering and design a "local-global" dual-branch Context Anchor Attention (CAA) algorithm integrated into the neck network. The local branch focuses on details, while the global branch models the equipment layout, enhancing the model's feature representation capabilities. Finally, to address the irregular shapes of underground PPE caused by occlusion, the neck network is replaced with a Dynamic Snake Convolution (DSConv). This reduces redundancy and enhances the model's ability to detect irregularly shaped personal protective equipment (PPE). The improved model achieves lightweightness while maintaining detection accuracy, achieving an ideal balance between speed and accuracy for PPE detection, providing strong support for safe coal mine production. Specifically, this method includes: Step 1: Data Collection Data collection is carried out by collecting on-site images with professional cameras and obtaining some public data sets from the Internet. The construction and processing of the miners' personal protective equipment data set are completed. The specific flow chart is as follows: Figure 1 shown.

[0028] Step 1.1: Build a high-quality dataset of 10,000 images. The dataset includes images from public datasets and images captured using professional cameras in various underground coal mine scenarios. The dataset covers different PPE wearing conditions and background conditions, and includes images under different lighting conditions, angles, distances, and viewing angles. This aims to enhance the diversity and richness of the training dataset.

[0029] Step 1.2: Detailed processing of the dataset. All collected images were manually labeled using the LabelImage image annotation tool, categorizing them into six main categories: hard hats, protective masks, mining lamps, and professional work clothes. After filtering and cleaning the labeled data, the annotated XML files were converted into VOC format files that can be recognized by the Yolo object detection algorithm.

[0030] Step 1.3: Divide the entire dataset into three parts: training set, validation set, and test set according to the ratio of 8:1:1.

[0031] Step 2: Build a target recognition model based on the YOLOv11 model Based on the YOLOv11 model, it is optimized and expanded to meet the special needs of underground miners' personal protective equipment (PPE) recognition. The improved model structure is as follows Figure 2 As shown in the figure, by introducing some new technical means such as more efficient feature extraction layer design, upper and lower anchor point attention to establish global semantic association, and snake-like dynamic convolution to adapt to target shape, the recognition challenges brought by the complex underground environment are effectively solved and the overall performance of the model is improved.

[0032] Step 2.1: Input an image of size 640×640×3 and send it to the backbone. In this application, the StarNet backbone network replaces the original YOLOv11 backbone network. The overall structure of StarNet is as follows: Figure 3 As shown in the figure, it enhances multi-scale feature extraction through star-shaped parallel branches and uses attention to focus on key areas, which is highly sensitive to small and occluded targets in complex scenes.

[0033] StarNet consists of four main stages, each of which uses modular components to gradually enhance feature representation while optimizing computational efficiency. The backbone network extracts image features and generates feature maps at different scales. The input image (640×640×3) is first convolved with a small kernel (3×3 kernel size, stride 2) to extract initial local features (edges, textures), significantly reducing the number of parameters. Batch normalization is then applied to normalize the feature distribution or rectified linear units to enhance numerical stability.

[0034] Specifically, in stage one, a depthwise separable convolution (DWConv) with a large kernel (7×7 kernel size, stride 1) is used to capture long-range spatial dependencies, compensating for the locality limitations of convolution with a small kernel. Element-wise multiplication is then used to fuse multi-scale features and enhance the saliency of the target region. Stages two and three perform further downsampling, multi-scale feature extraction, and spatial optimization. Stage four uses DWConv to capture global context for the final time, ensuring the integrity of high-level features. The output feature map is then combined with global average pooling or a fully connected layer.

[0035] Mining operations involve a variety of movements (such as bending, turning, and operating equipment) or obstruction by equipment, increasing the difficulty of detecting protective equipment (PPE). Leveraging its powerful feature extraction capabilities, the StarNet backbone network can effectively infer obscured PPE based on visible portions, maintaining recognition stability in different postures and significantly reducing missed detections. Furthermore, StarNet's multi-scale fusion capabilities enable it to accurately capture small targets (such as work boots and protective masks) in the harsh, low-light, dusty underground environment, further reducing the risk of missed detections. StarNet's lightweight nature, combined with YOLOv11's inherently efficient detection architecture, enables the model to run efficiently on underground edge devices, meeting the demands of real-time detection, enabling the high-precision, low-latency, real-time monitoring requirements of coal mines, and reducing hardware deployment costs.

[0036] Step 2.2: Output the feature map of the backbone network As the neck input, a “local-global” dual-branch CAA is designed and integrated into the neck network.

[0037] The dual-branch CAA structure includes a global branch and a local branch, such as Figure 4 As shown in the figure, local branches focus on details, while global branches model the device layout. The network structure of CAA is shown in the figure. Figure 5 As shown in Figure 2, it combines contextual anchors with dynamic attention weights, excelling at handling occlusions, small objects, and complex background interference. The high-resolution local branch focuses on capturing fine details of the miner and their equipment, while the low-resolution global branch reduces the feature map resolution through downsampling operations (such as pooling) to extract and understand global spatial context. This includes the miner's overall position and posture in the underground environment (such as bending over and walking direction), the relative positional relationships between different PPE (such as the spatial layout of hard hats, reflective vests, and mining lamps), and the interaction between the miner and surrounding background structures (such as tunnel walls and equipment outlines).

[0038] Specifically, such as Figure 4As shown in Figure 1, the input feature map is fed into two parallel processing branches. The high-resolution local branch maintains the high resolution of the original input and focuses on capturing fine details of the miner and his or her equipment (such as the reflective strips on the edge of the helmet, the outline of the work boots, the buckle of the self-rescuer, etc.). Dynamically generated anchor points are used to accurately locate these key local areas, which is crucial for recognizing small objects (such as mining lamps and protective masks) and partially occluded areas that often occur in complex underground environments. The low-resolution global branch reduces the resolution of the feature map through downsampling operations (such as pooling). Its core function is to extract and understand global spatial context information. This includes the miner's overall position and posture in the underground environment (such as bending over and walking direction), the relative positional relationships between different PPE (such as the spatial layout of the helmet, reflective vest, and mining lamp), and the interaction between the miner and surrounding background structures (such as the tunnel walls and equipment outlines). This global perspective helps the model stably infer the miner's overall presence and basic equipment configuration based on scene structure under harsh conditions such as low light and high dust. Finally, the two branches perform feature enhancement through local detail attention and global layout attention, respectively, and then fuse the features. Figure 5 As shown in Figure 1, CAA uses convolution kernels of different sizes in parallel. The small kernel branch, such as when n=0, has a kernel width of 11, capturing local details (such as edges and textures). As n increases, the kernel width expands to a larger range, and the large kernel branch extracts long-range contextual information (such as the overall structure of the object). Subsequently, weights are dynamically assigned through channel-spatial attention to enhance the target area features and suppress background noise. The input features are first average pooled to aggregate the global spatial information to generate low-resolution features. , where l represents the current layer (the lth layer), and l-1 represents the output of the previous layer as the input of the current layer. Subsequently, 1×1 convolution is used to compress the channel dimension to C / 2 to reduce the amount of computation: (1) Next, we use DWConv with a kernel size of 1×(11+2n) to capture long-distance dependencies in the horizontal direction and output features , apply DWConv with the same kernel size again, process the vertical context, and output features :

[0039] )(2) Finally, the horizontal and vertical features are concatenated or superimposed, and the channel dimension is restored to C through the second 1×1 convolution. The spatial attention weight is generated by the Sigmoid function to weight the original feature map, and the enhanced feature map is finally output. .

[0040] In summary, dual-branch CAA uses anchor points to accurately capture detailed features of protective equipment in dusty underground environments. The introduction of CAA enables the YOLOv11 neck network to dynamically perceive key contexts, significantly optimizing the performance of miner protective equipment detection. While ensuring real-time performance, it also improves the accuracy of small target recognition, providing more reliable technical support for mine safety monitoring.

[0041] Step 2.3: The enhanced feature map is further fed to the DSConv in the neck, whose structure is as follows Figure 6 As shown, DSConv is characterized by simulating the flexible path perception characteristics of snakes. It adapts to target geometry through dynamic paths, enhancing modeling capabilities for complex shapes while suppressing background noise interference. Furthermore, when changes in underground lighting or dust can cause feature distribution shifts, a conditional normalization layer is designed based on DSConv. This modulates the generation parameters of the dynamic convolution kernel based on real-time environmental characteristics, allowing the generation of the dynamic convolution kernel to better adapt to environmental conditions such as underground lighting and dust changes, thereby improving the performance and stability of the model in practical applications.

[0042] The structure of DSConv is as follows Figure 6 As shown in the figure, in DSConv, the feature map goes through the input feature map stage, the learning deformation stage, the snake convolution stage, the adaptive structural feature deformation stage, and the output stage. The core of the entire working sequence is to make the convolution operation better adapt to the structural characteristics of the input feature map through learning deformation and snake convolution, thereby extracting more useful features. This deformation convolution method may have more advantages than traditional convolution operations when processing image data with complex structures and diverse features, because it can dynamically adjust the shape and movement of the convolution kernel to capture more details and patterns. Specifically: First, in the input feature map stage, the input feature map contains the image data to be processed, and its format is usually a numerical matrix. Each value corresponds to a pixel value or feature value in the image, introducing a deformation offset. , a learned deformation function is applied to the input feature map to adjust its shape and orientation. The input feature map is fed into the learned deformation module, which learns deformation parameters based on the characteristics of the input feature map. The purpose of the learned deformation is to adapt to the structure and distribution of the input features and provide personalized deformation guidance for subsequent snake convolution operations.

[0043] During the snake convolution stage, the learned deformation is applied to the input feature map to perform a snake convolution operation. It should be noted that the DSConv in this application is equipped with a conditional normalization layer. The conditional normalization layer can obtain real-time environmental features and dynamically generate a dynamic convolution kernel. This dynamic convolution kernel is applied to the snake convolution, allowing the convolution operation to self-adjust according to actual environmental conditions. Through this dynamic adjustment mechanism, the convolution kernel can maintain a high degree of sensitivity to the characteristics of miners' personal protective equipment under different complex mine environmental conditions. At the same time, it effectively suppresses the interference of complex environmental background noise on feature extraction, improves the accuracy of feature extraction, enhances the real-time performance and reliability of the detection system, and provides a solid guarantee for the safe operation of miners.

[0044] The specific design and implementation of the conditional normalization layer are as follows: First, the downhole light intensity value E is obtained in real time through the light sensor light , use the dust sensor to measure the dust concentration in real time, and get the dust concentration estimate E dust Secondly, we design a fully connected conditional network to capture the real-time environment features E={E light ,E dust} as input and outputs conditional information for modulating the dynamic kernel generation parameters. After the introduction of the conditional normalization layer, the scaling factor γ∈R is generated by the conditional vector C. m and the offset factor β∈R m , perform affine transformation on the dynamic kernel generation parameters to obtain the dynamic kernel generation parameters θ after affine transformation cond , as shown in Equation (3). Then, the modulated dynamic kernel is used to generate the parameter θ cond , generates a dynamic convolution kernel K adapted to the current environmental conditions according to the input feature map, as shown in formula (4), where f gen Represents the dynamic kernel generation network. Finally, the generated dynamic convolution kernel K is used to perform a convolution operation on the input feature map X to obtain the output feature map O. The convolution operation formula is shown in Equation (5).

[0045] (3) (4) (5) In the snake convolution stage, the learned deformation is applied to the input feature map to perform a snake convolution operation. The process of snake convolution is as follows Figure 7As shown in the figure, DSConv is a 9×9 deformation two-dimensional transformation process. It cleverly integrates the idea of ​​dilated convolution and effectively expands the receptive field of the model by embedding a specific expansion rate in the standard convolution operation. This design not only enhances the model's ability to capture background information, but also cleverly avoids the problem of a sharp increase in computational cost caused by increasing the convolution kernel size. Let the coordinates of the standard convolution center be K i =(x i ,y i ), a convolution kernel K with a dilation rate of 1 can be expressed as formula (6): (6) Center grid K ij The position of x is the reference, i ±c, y i ±c and x j ±c, y j ±c represent the grid position K along the x and y axes respectively. i ±c, K j ±c is the horizontal and vertical coordinates, Δy represents the y-axis offset when changing along the x-axis, Δx represents the x-axis offset when changing along the y-axis, and c represents the horizontal offset of the center grid. The specific expansion process is as follows: (7) (8) During the adaptive structural feature warping stage, after the snake convolution operation, an intermediate feature map is obtained. This feature map contains the initially extracted feature information. Adaptive structural feature warping is then performed to further optimize the feature representation, making it more suitable for subsequent processing. Finally, the output feature map is obtained. This is the result of the snake convolution and adaptive structural feature warping, and contains richer and more discriminative feature information.

[0046] Because the model stochastically learns deformation offsets, the perception region may deviate from the target. Therefore, to ensure the continuity of the attention region throughout the entire processing process, DSConv employs an iterative strategy. This strategy sequentially matches each target to an observable location. This step-by-step approach ensures that even when processing irregular and complex PPE structures, it can adaptively focus on irregular local features based on structural morphological knowledge, preventing over-diffusion and loss of detail. Next, a dynamic convolution operation is performed on the input feature map using a set of learnable weights. This is achieved by varying the size, shape, or orientation of the weights. These weights can be dynamically adjusted based on the content of the input feature map to achieve better feature extraction. Finally, the offset at each location after the dynamic convolution is calculated, and the output feature map is generated based on the calculated offsets and the input feature map. In summary, DSConv has stronger feature representation capabilities for non-rigid targets in underground environments, such as the curved stripes on miners' reflective clothing and the curved edges of hard hats, reducing missed detections due to varying poses or occlusions. Secondly, its dynamic path focuses on local key areas, which can effectively distinguish protective equipment from similar backgrounds in complex scenarios, improve the positioning accuracy of small targets, and at the same time maintain the lightweight characteristics of the model to adapt to the real-time monitoring needs of mine edge equipment.

[0047] Step 3: Use the training set and validation set to train and validate the improved model.

[0048] Step 3.1: To ensure representativeness and reliability during model training and evaluation, we split the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1. The training set is used for the actual learning process. During training, the images in the dataset are preprocessed and resized to 640×640 pixels. The validation set is used to tune hyperparameters or select the best model version. The test set serves as a completely independent test of the performance of the selected model.

[0049] Step 3.2: To ensure fair and objective performance comparisons between different detection models, we adopted unified network training parameter settings. This aims to eliminate performance evaluation bias caused by parameter differences, allowing all models to demonstrate their true effectiveness under the same conditions. The number of training epochs (epochs) was set to 200, meaning the entire training dataset was repeatedly used 200 times to update the model weights. Furthermore, to process an appropriate amount of data in each iteration and maintain computational efficiency, the batch size (batch_size) was set to 8. A relatively high learning rate of 0.01 was used, which helps the model learn and adjust its weights quickly during the initial training phase. To ensure smoother transitions through local minima during the update process, the momentum was set to 0.937, a commonly used momentum value that accelerates convergence while maintaining model stability. Finally, to prevent overfitting and improve the generalization ability of the model, the weight decay is set to 0.0005. By adding a regularization term to the loss function to penalize large weight values, the model is encouraged to adopt simpler and more general assumptions and reduce sensitivity to noise and outliers in the training data.

[0050] Step 4: Use the trained model to identify the personal protective equipment of underground miners.

[0051] In summary, the improved YOLOv11, by incorporating a lightweight StarNet backbone, a local-global dual-branch CAA attention mechanism, and dynamic snake convolution (DSConv), achieves efficient recognition of protective equipment in complex underground environments. The model utilizes multi-scale feature fusion and adaptive weight adjustment to accurately detect critical equipment such as helmets and mining lamps while ensuring high real-time performance, providing a low-latency, highly robust intelligent monitoring solution for mine safety.

[0052] The recognition results before and after the model improvement are as follows Figure 8 As shown, Figure 8Insets (a1) and (a2) show the recognition results of the model before improvement, and insets (b1) and (b2) show the recognition results of the model before improvement. Compared with insets (a1) and (b1), the improved model has significantly improved the accuracy of target recognition, and the target frame calibration is more accurate. Compared with insets (a2) and (b2), the improved model has stronger anti-interference ability and can reduce the possibility of false detection. Mine environments are usually characterized by low light, high background complexity, and target occlusion, which place extremely high demands on target detection models. It can be clearly seen from the figure that the model in this paper has successfully overcome these challenges through targeted improvements. The YOLOv11 model has obvious false detection phenomena (misdetecting the equipment as a helmet) when facing insufficient light or complex background interference, which to a certain extent limits its application effect in complex environments. In contrast, the improved model proposed in this paper has shown significant improvement in these aspects, and can more accurately identify and locate personal protective equipment (PPE). Even in the complex and changing environment of mines, it can maintain high detection accuracy, showing a clear advantage in detecting PPE in complex scenarios.

[0053] This solution, based on the YOLOv11 model, combines StarNet, CAA, and DSConv to improve the backbone network and neck, respectively. This results in a lightweight PPE detection model based on the improved YOLOv11. StarNet is used to improve the backbone network, leveraging its attention mechanism on feature maps. This allows the model to automatically learn more discriminative features and remove unimportant information, thereby improving the model's feature extraction capabilities. CAA and DSConv are used to optimize the neck. A "local-global" dual-branch CAA is designed and integrated into the neck network. The attention mechanism adaptively allocates weights for different receptive fields, enabling the model to simultaneously focus on local details and global information, enhancing feature representation. DSConv adapts to target geometry through dynamic paths, enhancing its ability to model complex shapes while suppressing background noise. The design of the conditional normalization layer enables the generation of dynamic convolution kernels to better adapt to environmental conditions such as underground lighting and dust fluctuations, thereby improving the model's performance and stability in practical applications. The above technologies can ensure the improvement of detection accuracy and lightweight network in the complex and changeable coal mine environment, thereby improving the running speed and efficiency of the model.

[0054] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical spirit of this invention. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for identifying lightweight, irregular-shaped underground protective equipment, characterized in that: include, Step 1: Data collection: divide the collected data into different categories and divide the data set into training set, validation set, and test set; Step 2: Build a target recognition model to identify target features in the image data in the dataset. The target recognition model is based on the YOLOv11 model. Its backbone network is a star-shaped backbone network. The neck network is designed with dual-branch CAA and DSConv modules. The dual-branch CAA establishes global semantic associations through upper and lower anchor point attention. The DSConv module uses a dynamic kernel to perform snake-like dynamic convolution to adapt to the target shape. Step 3: Use the training set and validation set to train and validate the improved model; Step 4: Use the trained model to identify the personal protective equipment of underground miners.

2. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 1, characterized in that: In the step 2, the star-shaped backbone network is divided into four main stages. In the first stage, large-kernel depthwise separable convolution is used to capture long-distance spatial dependencies, make up for the local limitations of small-kernel convolution, and fuse multi-scale features using element-by-element multiplication operations; the second and third stages further downsample to perform multi-scale feature extraction and spatial optimization; the fourth stage captures the global context for the last time through DWConv to ensure the integrity of high-level features, and outputs feature maps in combination with global average pooling or fully connected layers.

3. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 2, characterized in that: In step 2, the dual-branch CAA includes parallel global and local branches. The high-resolution local branch focuses on capturing the fine details of the miners and their equipment, and the low-resolution global branch reduces the feature map resolution through downsampling operations to extract and understand global spatial context information.

4. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 3, characterized in that: In the step 2, the global branch and the local branch perform feature enhancement through local detail attention and global layout attention respectively, and then fuse the features.

5. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 3, characterized in that: The local branch generates low-resolution features , then use 1×1 convolution to compress the channel dimension to C / 2 to reduce the amount of calculation. Next, use DWConv with a kernel size of 1×(11+2n) to capture long-distance dependencies in the horizontal direction and output features , apply DWConv with the same kernel size again, process the vertical context, and output features Finally, the horizontal and vertical features are spliced ​​or superimposed, and the channel dimension is restored to C through the second 1×1 convolution. The spatial attention weight is generated by the Sigmoid function to weight the original feature map, and the enhanced feature map is finally output. , where l represents the current layer and l-1 represents the output of the previous layer as the input of the current layer.

6. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 1, characterized in that: In the step 2, the structure of the DSConv of the neck network includes an input feature map stage, a learning deformation stage, a snake convolution stage, an adaptive structural feature deformation stage and an output stage.

7. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 6, characterized in that: In the step 2, in the neck network, a conditional normalization layer is designed based on DSConv to modulate the generation parameters of the dynamic kernel according to the real-time environment characteristics, and the dynamic kernel is applied to the snake convolution.

8. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 7, characterized in that: The learning deformation stage learns deformation parameters according to the characteristics of the input feature map, provides personalized deformation guidance for the subsequent snake convolution operation, applies the learned deformation to the input feature map, and performs a snake convolution operation. After the snake convolution operation, an intermediate feature map is obtained, which contains the initially extracted feature information. Then, adaptive structural feature deformation is performed to further optimize the feature expression, and finally an output feature map is obtained.

9. The method for identifying lightweight underground irregular-shaped protective equipment according to claim 7, characterized in that: In the conditional normalization layer, first, the downhole light intensity value E is obtained in real time through the light sensor. light , use the dust sensor to measure the dust concentration in real time, and get the dust concentration estimate E dust ,Secondly, a fully connected conditional network is designed to take the real-time environment features E={E light ,E dust } as input and output the conditional information used to modulate the dynamic kernel generation parameters. After introducing the conditional normalization layer, the scaling factor γ∈R is generated by the conditional vector C. m and the offset factor β∈R m , perform affine transformation on the dynamic kernel generation parameters to obtain the dynamic kernel generation parameters θ after affine transformation cond , then, the modulated dynamic kernel is used to generate the parameters θ cond , according to the input feature map, a dynamic convolution kernel K that adapts to the current environmental conditions is generated. Finally, the generated dynamic convolution kernel K is used to perform a convolution operation on the input feature map X to obtain the output feature map O.

Images