A small target detection method for underground coal mine

Abstract

The application belongs to the technical field of target detection, and discloses a small target detection method for underground coal mines, which comprises the following steps: obtaining an image to be detected in an underground coal mine; inputting the image to be detected into a trained RPS-YOLO detection model; and outputting a detection result; wherein the RPS-YOLO detection model is improved based on a YOLOv11 network, and the improvement comprises the following steps: in the backbone network, replacing a C3K2 module with a DSCELAN module, and adding a first branch for outputting a first shallow feature map in a shallow layer of the network; in the neck network, replacing a C3K2 module of a feature pyramid structure with a DSCELAN module, and adding a P2 layer; in a fusion path of the P2 layer, adding a SOMA attention module; and in the detection head, adding a fine-grained detection head; the application fundamentally enhances the detail perception and semantic reservation capability for extremely small targets in underground coal mines while reducing the computational complexity, thereby effectively improving the detection precision and recall rate under the premise of ensuring real-time performance, and solving the problem of missed detection in a complex environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection, and particularly relates to a small target detection method for underground coal mines. BACKGROUND

[0002] In the field of intelligent construction and safety production monitoring of coal mines, it is of great significance to prevent safety accidents and ensure personnel safety by using visual technology to automatically detect whether underground workers wear safety helmets and carry self-rescuers and other key equipment. However, safety helmets, self-rescuers and other targets in the underground monitoring picture are usually very small objects (usually less than 32x32 pixels), which makes the detection task face serious challenges.

[0003] At present, for the small target detection in underground coal mines, the commonly used method is as follows: first, using a space-to-depth convolution module to extract features of the original image in the underground coal mine to retain more spatial details; second, introducing a self-attention module based on Transformer for global context modeling to suppress local overexposure interference; third, using a high-level screening feature pyramid combined with channel attention for multi-scale feature fusion; finally, predicting through a decoupled detection head and post-processing the prediction results.

[0004] However, the underground coal mine environment is complex, and the detection results are easily affected by light, dust and background. In addition, the posture of small targets is variable, and there may be partial occlusion. Therefore, the miss detection rate of this method for small targets in underground coal mines is still at a high level, and the ability to identify very small pixel size targets is limited. SUMMARY

[0005] The purpose of the present application is to provide a small target detection method for underground coal mines to improve the ability to identify small targets in underground coal mines.

[0006] The present application adopts the following technical solutions:

[0007] A small target detection method for underground coal mines, comprising the following steps:

[0008] Step 1: obtaining an image to be detected in an underground coal mine;

[0009] Step 2: inputting the image to be detected into a trained RPS-YOLO detection model;

[0010] Step 3: obtaining and outputting the target detection result of the RPS-YOLO detection model on the image to be detected;

[0011] The RPS-YOLO detection model is improved based on the YOLOv11 network, and the improvements include:

[0012] In the backbone network, the C3K2 module is replaced by a DSCELAN module, and a first branch for outputting a first shallow layer feature map is additionally arranged in the shallow layer of the backbone network;

[0013] In the neck network, the C3K2 module in the feature pyramid structure is replaced by a DSCELAN module; a P2 layer is additionally arranged in the feature pyramid structure to construct a multi-scale fusion path comprising the P2 layer, the P3 layer, the P4 layer and the P5 layer, wherein the input of the P2 layer is the first shallow layer feature map; in the fusion path of the P2 layer, a SOMA attention module is additionally arranged to weight the feature map fused in the P2 layer to obtain the feature map output by the P2 layer;

[0014] In the detection head, a fine-grained detection head connected with the feature map output by the P2 layer is additionally arranged.

[0015] The present application has the following advantages: the present application is improved based on YOLOv11, and the detection performance of small targets in the coal mine is significantly improved; the DSCELAN module is introduced in the backbone network and the neck network to enhance the detail feature extraction capability of irregular small targets; a high-resolution P2 layer is additionally arranged in the feature pyramid to construct a direct detection path for small targets, which effectively alleviates the problem of deep semantic loss; a SOMA attention module is embedded in the P2 layer fusion path to realize accurate focusing on the key area of small targets in a complex background; the three work together to make the final model achieve higher detection accuracy and recall rate for small size targets in the complex environment of the coal mine, while maintaining real-time inference capability, which significantly improves the detection accuracy and robustness in the coal mine. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The method flowchart of the present application is shown in the figure;

[0017] Figure 2 The principle diagram of the channel attention mechanism in the present application is shown in the figure;

[0018] Figure 3 The principle diagram of the spatial attention mechanism in the present application is shown in the figure;

[0019] Figure 4 The detection visualization result comparison diagram of different detection models in the present application and the prior art is shown in the figure;

[0020] Figure 5 The detection visualization result comparison diagram of the YOLOv11 detection model in the present application and the prior art is shown in the figure. DETAILED DESCRIPTION

[0021] The present application will be described in detail below with reference to the accompanying drawings and examples.

[0022] At present, an improved YOLO method for safety helmet wearing detection applied to underground coal mine scene is as follows:

[0023] (1) Feature extraction and detail enhancement: input the image into the improved CNN backbone network, which uses a spatial-to-depth convolution module to replace the original down-sampling layer. By interval sampling and channel splicing, the spatial resolution of the feature map is reduced while the detail information is transferred to the channel dimension, enhancing the ability to extract small target features in blurred images.

[0024] (2) Local overexposure suppression and feature interaction: input the features output by the end of the backbone network into the attention feature interaction module. Through the multi-head self-attention mechanism, the feature sequence is globally context modeled, enhancing the model's attention to the target area and suppressing the local overexposure interference caused by the mine lamp lighting.

[0025] (3) Multi-scale feature fusion optimization: a high-level screening feature fusion pyramid structure is used in the neck network. First, the channel attention mechanism is used to screen the importance of each scale feature, and then the selective feature fusion mechanism is used to weight and fuse the high-level semantic features and the screened low-level detail features, improving the model's detection ability for safety hats of different scales.

[0026] (4) Detection output and post-processing: input the fused multi-scale feature map into the decoupled detection head. Through convolution operation, the class probability and boundary box coordinates of each anchor point are predicted in parallel. Finally, the non-maximum suppression algorithm is used to remove redundant detection boxes, and the final safety hat detection result is output.

[0027] However, this method has the following shortcomings:

[0028] (1) The adaptability of the feature extraction mechanism to irregular small targets is insufficient: the spatial-to-depth convolution module used in this scheme effectively preserves spatial details, but its sampling pattern is fixed and regular. For small targets such as self-rescuers with irregular shapes and variable poses in underground coal mines, and safety hats partially occluded, it is difficult for this static convolution kernel to dynamically adjust its receptive field, limiting the modeling ability of target geometric deformation and the precision of feature extraction.

[0029] (2) The matching degree of the attention mechanism and the complex scene is not enough: the attention feature interaction module introduced in this scheme is based on Transformer self-attention, which can model global context, but in the scene with cluttered background and dense targets in underground coal mines, it is easily disturbed by a large amount of irrelevant background information.

[0030] (3) Multi-scale fusion strategy lacks specific attention to small targets: The high-level feature fusion pyramid structure performs feature selection through channel attention, but its fusion strategy is more general optimization. For small targets with weak responses in the feature map, their key information may be inadvertently diluted or filtered during multiple selection and fusion processes, resulting in insufficient semantic representation of small targets in deep feature maps.

[0031] To overcome the above shortcomings, the technical problems to be solved by the present application are:

[0032] (1) How to realize dynamic and fine feature extraction for irregular small targets.

[0033] (2) How to fundamentally strengthen the retention and transmission of small target information by the feature pyramid.

[0034] (3) How to design an attention mechanism that can specifically enhance the discriminative features of small targets in complex backgrounds.

[0035] The specific method of the present application is as follows:

[0036] In one embodiment, the present application provides a small target detection method for underground coal mines, which is particularly suitable for real-time and accurate identification of small targets such as safety helmets and self-rescuers with a size of less than 32x32 pixels.

[0037] As shown in Figure 1 , the method mainly includes the following steps:

[0038] Step 1: Obtain the image to be detected in the underground coal mine.

[0039] Step 2: Input the image to be detected into the trained RPS-YOLO detection model.

[0040] Step 3: Obtain and output the target detection result of the RPS-YOLO detection model on the image to be detected.

[0041] The RPS-YOLO detection model is improved based on the YOLOv11 network, and the improvements include:

[0042] First, in the backbone network, replace the C3K2 module with the DSCELAN module, and add a first branch for outputting the first shallow feature map in the shallow layer of the backbone network.

[0043] The method for extracting features by the DSCELAN module includes:

[0044] Channel compression is performed on the feature map of the image to be detected to obtain the initial projection feature.

[0045] Deep semantic features are extracted from the initial projection features by a reparameterization module.

[0046] The initial projection features are equally divided into two channel sub-features along the channel dimension.

[0047] The two channel sub-features are channel spliced with the deep semantic features to obtain a spliced feature map.

[0048] The spliced feature map is subjected to a depth convolution operation; the depth convolution operation specifically comprises: using a 3*3 convolution kernel to perform spatial convolution operation on each channel of the spliced feature map to extract spatial features and generate an intermediate feature map with unchanged channel number.

[0049] The intermediate feature map is subjected to a point-by-point convolution operation; the point-by-point convolution operation specifically comprises: using one or more 1*1 convolution kernels to linearly combine all channels of the intermediate feature map to fuse the channel information.

[0050] Specifically, the feature map of the image to be detected is input into a backbone network for feature extraction, the backbone network first performs channel compression on the feature map of the image to be detected to obtain initial projection features, and each layer of the backbone network comprises a DSCELAN module, the DSCELAN module is designed with two feature extraction paths, one path is used to keep the initial projection features unchanged and directly forward pass to retain the original feature information and avoid information loss caused by excessive transformation, and the other path is provided with a reparameterization module for extracting deep semantic features of the initial projection features.

[0051] The feature maps output from the above two feature extraction paths are spliced in the channel dimension to obtain a spliced feature map containing multi-source information, which provides rich and complementary input for subsequent feature fusion and enhancement.

[0052] Finally, the spliced feature map is subjected to a depth separable convolution process, and the feature map extracted by the backbone network is output, which can efficiently fuse multi-path information while significantly reducing the parameter quantity and computational complexity of the model.

[0053] The core of the DSCELAN module in the application is:

[0054] Multi-path design: a double-path parallel architecture is adopted, one path retains the details and textures of the input features through identity mapping or shallow processing to prevent the loss of spatial information of small targets, and the other path performs deep nonlinear transformation through a reparameterization module (such as RepNCSP) to learn high-level abstract features of small targets in the complex underground environment, the two paths provide complementary features of detail retention and semantic deepening.

[0055] The role of the reparameterization module: the module enhances the model expression ability and training stability in a multi-branch structure during training; during inference, it can be reparameterized into a single path, which retains the high performance gain obtained in the training phase under the premise of almost no increase in inference time delay and computational resource consumption, and adapts to the deployment requirements of downhole edge devices.

[0056] Efficient fusion mechanism: the deep separable convolution is used to fuse the two-way features, first through 3*3 convolution kernel to perform spatial convolution operation to extract spatial features of each channel, and then through 1*1 convolution kernel to perform cross-channel information fusion, which significantly reduces the calculation cost, realizes the effective aggregation of multi-path features, and guarantees the lightweight of the overall scheme.

[0057] Therefore, the DSCELAN module of the present application solves the problems of easy loss of small target details, difficult semantic discrimination and limited computing power in downhole.

[0058] Secondly, in the neck network, the C3K2 module in the feature pyramid structure is replaced by the DSCELAN module; a P2 layer is added in the feature pyramid structure to construct a multi-scale fusion path comprising the P2 layer, the P3 layer, the P4 layer and the P5 layer, wherein the input of the P2 layer is the first shallow layer feature map; in the fusion path of the P2 layer, a SOMA attention module is added for weighting the feature map fused in the P2 layer to obtain the feature map output by the P2 layer.

[0059] The feature pyramid network of the present application adds a P2 layer with higher spatial resolution to the YOLOv11 (multi-scale fusion path comprising the P3 layer, the P4 layer and the P5 layer), thereby constructing a multi-scale fusion path comprising the P2 layer, the P3 layer, the P4 layer and the P5 layer.

[0060] The fusion path of the P2 layer can output a feature map with the highest spatial resolution and a downsampling multiple of 4, which can retain extremely rich details, explicitly strengthen the network's ability to model small targets in space, and effectively improve the problem that small targets are weakly responsive and easily lost in high-level semantic feature maps.

[0061] The SOMA attention module is improved based on the Convolutional Block Attention Module (CBAM), and the improvements include:

[0062] In the channel attention mechanism, a first channel convolution layer, a GELU activation function and a second channel convolution layer are sequentially connected to replace the fully connected layer in the CBAM attention module to generate channel attention weights; the convolution kernels of the first channel convolution layer and the second channel convolution layer are both 1*1.

[0063] In the spatial attention mechanism, a parallel first spatial convolution layer and a second spatial convolution layer are adopted to replace the 7x7 convolution layer in the CBAM attention module to extract and fuse multi-scale spatial features to generate spatial attention weights; the convolution kernel of the first spatial convolution layer is 3x3, and the convolution kernel of the second spatial convolution layer is 7x7.

[0064] The method for generating channel attention weights by the channel attention mechanism includes the following steps:

[0065] The feature map fused in the P2 layer is subjected to global average pooling to generate a channel feature vector.

[0066] The channel feature vector is sequentially subjected to dimension reduction by a first channel convolution layer, nonlinear transformation by a GELU activation function, and dimension increase by a second channel convolution layer.

[0067] The output after dimension increase is subjected to Sigmoid activation to generate channel attention weights.

[0068] As shown in Figure 2 , the specific process of the channel attention mechanism is as follows:

[0069] The feature map fused in the P2 layer is subjected to global average pooling to aggregate information of all spatial positions in each channel to obtain a one-dimensional channel feature vector.

[0070] The channel feature vector is input into a multi-layer perceptron for processing. In the multi-layer perceptron, a first channel convolution layer with a convolution kernel of 1x1 is first used to reduce the dimension of the channel feature vector to extract more discriminative compact channel features under low illumination and dust interference, thereby reducing parameters and preventing overfitting. Then, a GELU activation function is used to nonlinearly transform the channel feature vector after dimension reduction to provide smoother gradient flow and enhance the representation ability of small targets. Finally, a second channel convolution layer with a convolution kernel of 1x1 is used to increase the dimension of the nonlinearly transformed channel feature vector to restore the original channel dimension.

[0071] The dimension-increased channel feature vector output by the multi-layer perceptron is normalized by a Sigmoid function to obtain channel attention weights. The channel attention weights are multiplied with the feature map fused in the P2 layer channel by channel to obtain a feature map weighted by the channel attention mechanism, wherein represents channel-by-channel multiplication.

[0072] The method for generating spatial attention weights by the spatial attention mechanism includes the following steps:

[0073] The feature map weighted by the channel attention weights is subjected to global average pooling and global maximum pooling respectively to obtain channel mean features and channel maximum features.

[0074] The channel mean feature and the channel maximum value feature are spliced in the channel dimension to obtain an aggregated feature map.

[0075] The aggregated feature map is input into a first spatial convolution layer and a second spatial convolution layer to extract local scale space features and global scale space features respectively.

[0076] The local scale space features are fused with the global scale space features, and batch normalization processing is performed on the fused features.

[0077] Sigmoid activation is performed on the output of the batch normalization processing to generate spatial attention weights.

[0078] As shown in Figure 3 , the specific process of the spatial attention mechanism is as follows:

[0079] The feature map weighted by the channel attention mechanism is subjected to global average pooling and global maximum pooling respectively.

[0080] The results of the two kinds of pooling are spliced in the channel dimension to obtain an aggregated feature map, which aggregates the statistical information of each spatial position in the channel dimension.

[0081] A first spatial convolution layer with a convolution kernel of 3x3 and a second spatial convolution layer with a convolution kernel of 7x7 are used in parallel to process the aggregated feature map to capture spatial context of different scales at the same time.

[0082] The features output by the two parallel convolution branches are fused, and a batch normalization operation and a Sigmoid function are used to generate spatial attention weights.

[0083] The generated spatial attention weights are multiplied pixel by pixel with the feature map weighted by the channel attention mechanism to obtain the feature map output by the P2 layer, wherein represents a fusion operation, represents a channel-by-channel multiplication.

[0084] The SOMA attention module of the present application is an improvement on the attention module represented by CBAM in view of the shortcomings exposed in the complex scene of coal mines.

[0085] In the channel attention mechanism of the SOMA attention module:

[0086] Firstly, only global average pooling is used for the fused feature map of P2 layer, and the global maximum pooling in CBAM is omitted. In the complex scene of coal mine underground, local overexposure (such as strong light of miner's lamp) and random noise are common. Global maximum pooling is easily disturbed by such extreme values, thus amplifying the interference signal of irrelevant background. By removing the global maximum pooling branch, on the one hand, the computational complexity is reduced, and the processing efficiency of SOMA attention module is improved. On the other hand, the model becomes more dependent on global average statistics, enhancing the stable perception of the overall feature distribution of small targets, and avoiding over-sensitive response to local noise. This improvement makes the channel attention mechanism more suitable for the characteristics of small targets in coal mine underground, such as weak features and strong background interference, thus improving the robustness and accuracy of feature calibration.

[0087] Secondly, a multi-layer perceptron structure including a first channel convolutional layer, a GELU activation function and a second channel convolutional layer is adopted. The channel attention mechanism of existing CBAM module usually adopts a fully connected layer structure. When performing global mapping in the channel dimension, this structure tends to uniformize the information of all spatial positions. For small targets in underground (such as a self-rescuer far away), whose feature response is already weak, their key local detail features may be diluted in the global average process, or even suppressed by background noise. The multi-layer perceptron structure adopted by the present application can better capture and enhance small target features (such as the edge of a safety helmet) that appear in a specific local area, avoiding the destruction of spatial structure by the fully connected layer. The GELU activation function provides a smoother nonlinear transformation, which helps the network to model the complex relationships between small target fine-grained features more finely, thus significantly improving the amplification ability of weak but discriminative channel features.

[0088] In the spatial attention mechanism of the SOMA attention module:

[0089] The first spatial convolutional layer and the second spatial convolutional layer are used in parallel to process the aggregated feature map, and then the outputs of the two are fused, and finally the spatial attention weight is generated through batch normalization. The spatial attention mechanism of CBAM generates spatial attention weight through a single 7x7 convolution. This large receptive field and single scale design works well in general datasets with relatively simple backgrounds, but it is difficult to balance in coal mine underground where there are extreme clutter and interference sources of different scales (such as coexistence of metal reflection near and dust interference far away). The first spatial convolutional layer of the present application is responsible for capturing local detail features such as fine edges and corner points of small targets. The second spatial convolutional layer is responsible for modeling the global context relationship between the target and the surrounding roadway and equipment, which helps to understand the reasonableness of the target's existence, thus suppressing irrelevant background interference in a large range. Batch normalization operation effectively stabilizes the training process of attention weight, and enhances the generalization ability of the model on high-noise and non-uniform illumination data in underground.

[0090] Therefore, the SOMA attention module of the present application has the core advantage of specially solving the problem of weak small target signal and strong background interference in the coal mine underground, and the method enhances the subtle features of the small target and suppresses the complex background noise and irrelevant interference in the underground through the cooperative work of the channel attention and the spatial attention, thereby improving the positioning accuracy and recognition robustness of the model to the small target in the extreme environment.

[0091] Finally, a fine-grained detection head connected with the feature map output by the P2 layer is added in the detection head.

[0092] The P2 layer in the feature pyramid structure is responsible for generating a high-resolution and detail-rich feature map, and the newly added fine-grained detection head is specially used for analyzing the feature map to complete the classification and positioning prediction of the target. This design solves the problem that the original detection head in the YOLOv11 network cannot effectively utilize the shallow layer detail information due to the lack of corresponding high-resolution input. The fine-grained detection head enables the network to directly perceive and locate the small target at the level closest to the original pixel information, effectively avoiding the semantic dilution and spatial information loss caused by transmitting the small target information to the deep layer for detection. It, together with the first branch in the backbone network for outputting the first shallow layer feature map, the P2 layer of the feature pyramid structure and the SOMA attention module, forms a special path for the detection of small targets in the coal mine underground, thereby structurally strengthening the detection capability of the model to small targets and being the key design for significantly improving the recall rate and reducing the small target missing rate.

[0093] The present application will be described in detail below in combination with simulation experiments.

[0094] (1) Technical effect verification data: The present application has been fully experimentally verified on the self-built coal mine underground safety helmet and self-rescuer detection data set, and the key quantitative effect comparison is as follows.

[0095] (2) Ablation experiment: The present application is subjected to an ablation experiment on the YOLOv11 model, and the experimental result data is shown in Table 1. In the table, A is the DSCELAN module, B is the fine-grained detection head, C is the SOMA attention module, P is the average precision, R is the recall rate, mAP50 is the average precision when the threshold is 0.5, mAP0.5:0.95 is the comprehensive average precision, and FPS is the inference speed. As can be seen from the table, the effect of each of the three modules used alone is better than that of YOLOv11, and therefore, through the ablation experiment, it can be verified that the DSCELAN module, the fine-grained detection head and the SOMA attention module proposed in the present application all make independent and collaborative contributions to the performance improvement, and when the three modules work together, the model reaches the optimal performance.

[0096] Table 1 Ablation experiment results

[0097]

[0098] (3) Compared with the baseline model: YOLOv11n is selected as the baseline model, and the comparison results are shown in Table 2. In the table, P is the average precision, R is the recall rate, mAP50 is the average precision when the threshold is 0.5, and FPS is the inference speed. It can be seen from the table that the model (RPS-YOLO) of the present application has a significant improvement in key evaluation indicators compared with YOLOv11n: mAP50 increases by 5.3%, R increases by 6.8%, which directly proves the effectiveness of the present application in the task of small target detection in coal mine underground.

[0099] (4) To further verify the effectiveness of the present application in the small target detection scene in coal mine underground, the present application is compared with the current mainstream lightweight target detection models YOLOv5s, YOLOv8n, YOLOv10n, YOLOv11n, YOLOv12n, RT-DETR (ZHAO Y, LV W, XU S, et al. Detrs beat yolos on real-time object detection [C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024: 16965-16974.) and the method in document 1 (ZOU C, YU S Q, YU Y K, et al. Side-Scan Sonar Small Objects Detection Based on Improved YOLOv11 [J]. Journal of Marine Science and Engineering, 2025, 13(1): 162.) Systematic comparative experiments are carried out, and the experimental results are shown in Table 2. It can be seen from Table 2 that the present application achieves the optimal performance in the three key indicators of P, R and mAP50, among which P reaches 91.3%, which is 5.6%, 5.8% and 3.9% higher than YOLOv5s, YOLOv8n and YOLOv11n respectively; R increases to 76.4%, which is significantly better than the 69.8% of YOLOv10n and the 69.6% of YOLOv11n; mAP50 reaches 83.8%, so the performance of the model of the present application is higher than that of other models in Table 2, and the FPS of the present application is kept at 185.0 while maintaining high accuracy, although it is slightly lower than the models such as YOLOv10n and YOLOv5s, but still meets the current industrial real-time requirements, indicating that the present application realizes significant improvement in accuracy while ensuring real-time performance.

[0100] Table 2 Comparison of detection results of different models

[0101]

[0102] Figure 4 For comparison of detection visualization results of different detection models in the present application and prior art, in the figure, helmet is a safety helmet, self rescuer is a self-rescuer, and the numbers represent confidence, wherein, scene one and scene two are located in the mining working face area, and there are typical environmental disturbances such as uneven brightness distribution, strong light lamp irradiation at the top and complex metal structure in this area; scene three is located in the mine transportation roadway area, the overall illumination of this scene is very low, only relying on the weak light source of the roadway top or the equipment itself, the contrast is low, and it is easy to cause model missed detection or confidence decrease.

[0103] As can be seen from the figure, in scene one, the YOLOv8n model only detects the safety helmet, and does not detect the self-rescuer, which appears to be missed detection; the YOLOv10n model and the YOLOv11n model identify the high-reflective components such as pipeline valves and metal connectors as self-rescuers, and do not detect the safety helmet, which appears to be false detection and missed detection; the method of the present application accurately detects the safety helmet and the self-rescuer, and the confidence is higher than the detection results of other existing models.

[0104] In scene two, the YOLOv8n model only detects the safety helmet, and does not detect the self-rescuer, which appears to be missed detection; the YOLOv10n model detects the safety helmet and the self-rescuer, but the confidence is lower than the method of the present application; the YOLOv11n model does not detect the safety helmet and the self-rescuer, which appears to be missed detection; the method of the present application accurately detects the safety helmet and the self-rescuer, and the confidence is higher than the detection results of other existing models.

[0105] In scene three, the YOLOv8n model detects two safety helmets and one self-rescuer, but the confidence is lower than the method of the present application; the YOLOv10n model detects one safety helmet and one self-rescuer, and fails to detect all safety helmets, which appears to be missed detection; the YOLOv11n model detects two safety helmets, and fails to detect the self-rescuer, which appears to be missed detection; the method of the present application accurately detects two safety helmets and one self-rescuer, and the confidence is higher than the detection results of other existing models.

[0106] Therefore, it can be known that the method of the present application effectively suppresses the false detection problem caused by strong light reflection in the case of strong light irradiation, successfully distinguishes the metal components from the self-rescuer, and can accurately detect the safety helmet and the self-rescuer and other key targets, and the detection result is more stable and the confidence is higher; in the case of very low light, the false judgment problem caused by shadow and background noise is effectively reduced, and stronger low-light robustness and complex background separation ability are embodied.

[0107] Figure 5 This is a comparison chart of the detection visualization results of the present invention and the existing YOLOv11 detection model; in the chart, helmet is a safety helmet and self rescuer is a self-rescuer.

[0108] Scenario 4 is located in the mining face area. The top of the mining face is strongly illuminated by overhead lights, resulting in obvious overexposed areas. The complex structure of the equipment on site, such as hydraulic supports and metal pipelines, creates highly bright and highly reflective areas, which can easily interfere with detection. As can be seen from the heat map of the YOLOv11 detection results, the activated areas are clearly biased towards metal parts, and the hot spots are diffuse. The YOLOv11 detection model mistakenly identifies the metal reflection area as a feature of the self-rescue device, resulting in false detection. As can be seen from the heat map of the detection results of this invention, the heat map shows a clear and high response to the safety helmet and self-rescue device, and is almost unaffected by metal reflection.

[0109] Scene 5 is located in a mine tunnel area, where there is significant backlighting. The main light source is LED strips, and the subjects are mainly in backlight, resulting in blurred outlines and low contrast. The heatmap from YOLOv11 shows a wide activation range, with some background walls and track areas also activated. The confidence area is scattered, and multiple small targets have unstable responses. The heatmap from this invention shows that the activation points are close to the safety helmets and self-rescue devices, and the focus remains even when multiple people are detected simultaneously, with almost no invalid activations in the background.

[0110] Scene 6 is located in a pedestrian passageway (with localized strong light). The lighting in the passageway is relatively concentrated, creating a semi-backlight effect when approaching the light source. The heatmap from the YOLOv11 detection results shows that the activation area is relatively diffuse, with some responses falling on the passageway walls and signs, resulting in insufficient focus on people and unstable detection results. The heatmap from the detection results of this invention shows that the heatmap is concentrated in the area of ​​the safety helmet and the back self-rescue device, maintaining good differentiation for multiple targets and accurately locating targets at the boundary between light and dark.

[0111] Scene 7 is located in the tunneling return airway (low illumination + dust interference + dense personnel) area. The area is dimly lit with low overall brightness and some dust, resulting in heavy noise in the image. It is a typical complex scene. The heat map of the YOLOv11 detection results shows that the activation points are scattered and the response to the safety helmets is weak, resulting in missed detections. The heat map of the detection results of this invention shows that the safety helmets and self-rescue devices of multiple personnel show obvious high activation. Even in dark areas, the model can still identify key parts.

[0112] (5) In this experiment, the Ubuntu 22.04 operating system is used, based on the PyTorch2.5.1 deep learning framework and the Python 3.12 environment, using CUDA 12.4 for model training and testing. The specific experimental environment and training parameters are shown in Table 3 and Table 4.

[0113] (6) The data set used in this experiment: The data set used in the experiment is collected by deploying high-definition cameras in multiple underground coal mine scenes, including coal mining faces, transportation tunnels, and underground stations, to ensure that diverse scenes and activities can be captured.

[0114] Table 3 Experimental environment configuration

[0115]

[0116] Table 4 Experimental model parameter table

[0117]

Claims

1. A method for detecting small targets in underground coal mines, characterized in that, Includes the following steps: Step 1: Acquire the image to be inspected in the coal mine; Step 2: Input the image to be detected into the trained RPS-YOLO detection model; Step 3: Obtain and output the target detection results of the RPS-YOLO detection model in the image to be detected; The RPS-YOLO detection model is an improvement based on the YOLOv11 network, and the improvements include: In the backbone network, the C3K2 module is replaced with the DSCELAN module, and a first branch for outputting the first shallow feature map is added to the shallow layer of the backbone network. In the neck network, the C3K2 module in the feature pyramid structure is replaced with the DSCELAN module; a P2 layer is added to the feature pyramid structure to construct a multi-scale fusion path containing layers P2, P3, P4, and P5, wherein the input of the P2 layer is the first shallow feature map; in the fusion path of the P2 layer, a SOMA attention module is added to weight the feature map after fusion of the P2 layer to obtain the feature map output by the P2 layer. A fine-grained detection head connected to the feature map output by the P2 layer is added to the detection head; The method for extracting features using the DSCELAN module includes: Channel compression is performed on the feature map of the image to be detected to obtain initial projection features; Deep semantic features are obtained by extracting deep semantic information from the initial projection features through the reparameterization module; The initial projection feature is equally divided into two channel sub-features along the channel dimension; The two channel sub-features and deep semantic features are concatenated to obtain a concatenated feature map, and the concatenated feature map is then subjected to convolutional fusion processing to output an initial feature map. The SOMA attention module is an improvement based on the CBAM attention module, and the improvements include: In the channel attention mechanism, a first channel convolutional layer, a GELU activation function, and a second channel convolutional layer are sequentially connected to replace the fully connected layer in the CBAM attention module to generate channel attention weights; the convolutional kernels of the first and second channel convolutional layers are both 1×1. In the spatial attention mechanism, a parallel first spatial convolutional layer and a second spatial convolutional layer are used to replace the 7×7 convolutional layer in the CBAM attention module, so as to extract and fuse multi-scale spatial features to generate spatial attention weights; the convolutional kernel of the first spatial convolutional layer is 3×3, and the convolutional kernel of the second spatial convolutional layer is 7×7.

2. The method for detecting small targets in underground coal mines according to claim 1, characterized in that, The convolutional fusion process for the spliced ​​feature maps includes: A depthwise convolution operation is performed on the stitched feature map; specifically, the depthwise convolution operation is as follows: a spatial convolution operation is performed on each channel of the stitched feature map using a 3×3 convolution kernel to extract spatial features and generate an intermediate feature map with the same number of channels; The intermediate feature map is subjected to a pointwise convolution operation; the pointwise convolution operation specifically involves using one or more 1×1 convolution kernels to linearly combine all channels of the intermediate feature map to fuse the information of each channel.

3. The method for detecting small targets in underground coal mines according to claim 1, characterized in that, The feature map output by the P2 layer is downsampled by a factor of 4.

4. The method for detecting small targets in underground coal mines according to claim 1, characterized in that, The method for generating channel attention weights using the channel attention mechanism includes the following steps: Global average pooling is performed on the feature map after fusion at the P2 layer to generate channel feature vectors; The channel feature vectors are sequentially passed through a first channel convolutional layer for dimensionality reduction, a GELU activation function for nonlinear transformation, and a second channel convolutional layer for dimensionality increase. The output after dimensionality upscaling is activated by Sigmoid to generate channel attention weights.

5. The method for detecting small targets in underground coal mines according to claim 4, characterized in that, The method for generating spatial attention weights using the spatial attention mechanism includes the following steps: The feature map after being weighted by the channel attention weights is subjected to global average pooling and global max pooling respectively to obtain the channel mean feature and the channel maximum feature. The channel mean feature and the channel maximum feature are concatenated along the channel dimension to obtain an aggregated feature map; The aggregated feature map is simultaneously input into the first spatial convolutional layer and the second spatial convolutional layer to extract local scale spatial features and global scale spatial features, respectively. The local scale spatial features are fused with the global scale spatial features, and the fused features are then subjected to batch normalization. Sigmoid activation is applied to the output after batch normalization to generate spatial attention weights.

6. The method for detecting small targets in underground coal mines according to claim 5, characterized in that, The method for obtaining the feature map output by the P2 layer is as follows: The generated spatial attention weights are multiplied pixel-by-pixel with the feature map weighted by the channel attention mechanism to obtain the feature map output by layer P2.

Images