A lightweight small target detection method for complex scenes of coal mines
By using an improved ML-YOLO model and employing C3k2_MDConv, LAPConv, and MBPA detection heads, the problems of lightweight models and insufficient accuracy in small target detection in underground coal mines were solved, achieving efficient and accurate detection in complex scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN UNIV OF SCI & TECH
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies for detecting small targets in underground coal mines suffer from increased model parameters and computational overhead, making it difficult to meet the lightweight requirements of edge computing scenarios. At the same time, the detection accuracy is insufficient, especially in complex scenarios where small targets are easily lost or falsely detected.
An improved ML-YOLO model is adopted. By replacing the convolutional modules of the YOLOv11s model with C3k2_MDConv and LAPConv modules in the backbone network, and using the MBPA detection head in the detection head, the feature extraction and detection process is optimized, the number of model parameters and computational overhead are reduced, and the detection accuracy of small objects is improved.
It significantly reduces the number of model parameters and computational overhead, improves the detection accuracy and robustness of small targets in complex downhole scenarios, and achieves reliable detection in resource-constrained environments.
Smart Images
Figure CN122289653A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target detection technology, specifically relating to a lightweight small target detection method for complex coal mine scenarios. Background Technology
[0002] In the field of intelligent construction and safety monitoring in coal mines, the automatic identification of whether underground personnel are wearing safety helmets and carrying self-rescue devices as required by regulations through visual technology is crucial for preventing accidents and ensuring safety. However, these targets are often extremely small in size (usually less than 32×32 pixels) in the monitoring screen, posing a severe challenge to the detection work.
[0003] In existing technologies, the small target detection algorithm for downhole safety helmets based on REIW-YOLOv10n improves feature extraction capabilities by designing the RepNMSC module to replace the original C2fUIB structure, introducing reparameterization technology and normalized attention mechanism; it also replaces the original PANet with the ERepGFPN structure to optimize the feature pyramid topology; and adds a P2 small target detection head to improve the accuracy of small target detection.
[0004] However, the essence of this method is to improve accuracy by increasing the number of detection head layers, which directly leads to an increase in the number of model parameters and computational overhead, which contradicts the core requirement of lightweight models in edge computing scenarios (such as real-time monitoring of underground coal mines). Summary of the Invention
[0005] The purpose of this invention is to provide a lightweight small target detection method for complex coal mine scenarios, so as to meet the deployment requirements of lightweight models in edge computing scenarios while significantly reducing the number of model parameters and computational load.
[0006] The present invention adopts the following technical solution: A lightweight small target detection method for complex coal mine scenarios includes the following steps: The image to be detected is input into the trained ML-YOLO model; Obtain and output the target detection results of the ML-YOLO model in the image to be detected; The ML-YOLO model is an improvement upon the YOLOv11s model. The ML-YOLO model includes a backbone network, a neck network, and a detection head. Improvements include: In the backbone network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module, and the convolutional module in the deep extraction stage of the YOLOv11s model is replaced with the LAPConv module. In the neck network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module; In the detection head, the original detection head of the YOLOv11s model is replaced with the MBPA detection head.
[0007] The present invention also provides a lightweight small target detection device for complex coal mine scenarios, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements any of the above methods when executing the computer program.
[0008] The beneficial effects of this invention are as follows: By replacing the C3k2_MDConv and LAPConv modules in the backbone network, replacing the neck network with the C3k2_MDConv module, and replacing the detection head with the MBPA detection head, this invention abandons the existing architectural approach of improving accuracy by adding detection head layers. The lightweight design of modulated deformable convolution in the C3k2_MDConv module, the channel redundancy reduction mechanism of partial channel convolution in the LAPConv module, and the parameter optimization design of the dual-path heterogeneous structure in the MBPA detection head enable the overall model to significantly reduce the number of parameters and computational overhead while maintaining detection capabilities. Therefore, this invention has better lightweight performance than the prior art and can better adapt to the requirements of model deployment in the edge computing scenario of underground coal mines. Attached Figure Description
[0009] Figure 1 This is a network architecture diagram of the ML-YOLO model in this invention; Figure 2 This is a network architecture diagram of the C3k2_MDConv module in this invention; Figure 3 This is a network architecture diagram of the LAPConv module in this invention; Figure 4 This is a network architecture diagram of the MBPA detection head in this invention; Figure 5 This is a comparison chart of the detection visualization results of the present invention and different detection models in the prior art. Detailed Implementation
[0010] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0011] In one embodiment, the present invention provides a lightweight small target detection method for complex coal mine scenarios. This method can reduce the number of model parameters and computational load while ensuring real-time and accurate identification of small targets with a scale of less than 32×32 pixels, such as safety helmets and self-rescue devices.
[0012] This method mainly includes the following steps: The image to be detected is input into the trained ML-YOLO model; Obtain and output the target detection results of the ML-YOLO model in the image to be detected.
[0013] like Figure 1 As shown, the ML-YOLO model is an improvement upon the YOLOv11s model. The ML-YOLO model includes a backbone network, a neck network, and a detection head. The improvements include: In the backbone network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module, and the convolutional module in the deep extraction stage of the YOLOv11s model is replaced with the Lightweight Adaptive Partial Convolution (LAPConv) module.
[0014] In the neck network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module.
[0015] In the detection head, the original detection head of the YOLOv11s model is replaced with a dual-path heterogeneous detection head based on Multi-Branch Parallel Attention (MBPA).
[0016] In one embodiment, the C3k2_MDConv module is an improvement based on the C3k2 module, and the improvements include: Replace the Bottleneck unit in the C3k2 module with the Bottleneck_MDConv unit.
[0017] The Bottleneck_MDConv unit is obtained by replacing the standard convolution in the Bottleneck unit with Modulated Deformable Convolution (MDCN).
[0018] like Figure 2 As shown, the main body of the C3k2_MDConv module is an improved C2f architecture framework. Its internal core consists of multiple stacked Bottleneck_MDConv units, each of which replaces the standard convolution with MDCN.
[0019] MDCN dynamically adjusts the sampling position of the convolution kernel through learnable spatial offsets, and introduces learnable feature modulation weights for each sampling point. After nonlinear mapping by the Sigmoid function, the contribution of each sampling point is dynamically weighted. This design enables the convolution kernel to adaptively adapt to small targets such as downhole safety helmets and self-rescue devices based on input features without significantly increasing computational complexity. At the same time, it effectively suppresses the interference of invalid features introduced when the sampling point shifts to the background area. It significantly improves the ability to extract detailed features such as edges and textures of small targets and the deformation adaptability from the source of the backbone network.
[0020] In the feature stream, the input features are segmented into multiple branches through the C2f framework. Some branches undergo adaptive feature transformation via the Bottleneck_MDConv unit, while others retain the original information through shortcut connections. Finally, the features of all branches are concatenated and fused along the channel dimension.
[0021] Specifically, the following steps are taken: First, a lightweight feature extraction unit, Bottleneck_MDConv, is constructed. This module takes a standard convolutional layer as input and enhances the model's feature perception ability for small targets in complex lighting and heterogeneous backgrounds in coal mines by introducing MDCN. This effectively improves the diversity and stability of features without significantly increasing computational complexity. On this basis, a mid-level feature fusion module, C3kMDConv, is constructed. In the traditional C3 structure, Bottleneck_MDConv replaces the standard Bottleneck, enabling embedding in the mid-level feature extraction stage, thereby improving the network's feature separation ability in complex backgrounds. Finally, combined with the lightweight C2f framework, a high-level module, C3k2_MDConv, is proposed. In the feature flow, multi-level deformable convolutional units are used to achieve adaptive feature fusion and cross-scale transfer. This significantly reduces the number of model parameters while improving the diversity and stability of feature expression. This fundamentally improves the model's adaptability to geometric deformation and detail discrimination ability for small targets (such as safety helmets and self-rescue devices) in complex coal mine scenarios, while maintaining the model's lightweight characteristics.
[0022] In one embodiment, the feature processing in the LAPConv module includes: The input feature map is adaptively weighted and fused in the spatial dimension by using lightweight adaptive attention, and then purified in the channel dimension by partial channel convolution to obtain the enhanced feature map.
[0023] like Figure 3As shown, the LAPConv module includes a Lightweight Adaptive Enhancement (LAE) module and a Partial Convolution (PConv) module connected in sequence.
[0024] The feature processing in the LAE module includes: The input feature map is subjected to average pooling, and contextual information is extracted through 1×1 convolution to obtain intermediate features; The intermediate features are reorganized from the spatial dimension to the block dimension through a recombination operation; The recombined features are normalized using the Softmax function to generate a spatial attention weight map. Grouped downsampling convolution is performed on the input feature map to extract multi-scale features; The spatial attention weight map and multi-scale features are summed element-wise to obtain the spatial augmented features.
[0025] The feature processing in the PConv module includes: Spatial augmentation features are divided into convolutional branches and identity mapping branches according to channels; Perform convolution transformation on the convolution branch; The transformed convolutional branch and the identity mapping branch are concatenated and fused through channels to obtain intermediate fused features; The intermediate fusion features are mixed using a projection layer to combine channel information, and the final enhanced feature map is output by combining residual connections.
[0026] The convolution transformation of the convolution branch includes: performing 3×3 depth convolution, batch normalization, and ReLU activation on the convolution branch.
[0027] Specifically: In the LAE module, firstly, average pooling is performed on the input feature map to capture contextual information. Then, a 1×1 convolution is used to map it to the original attention logic. To ensure the smooth progress of the reorganization operation, when the spatial size of the input feature map is not even, an edge padding strategy is used for preprocessing. Finally, a normalized spatial attention weight map is obtained through the reorganization operation and the Softmax function. The reorganization operation reorganizes the feature map from the spatial dimension to the block dimension, thereby generating four independent attention weights for each 2×2 local region. The input features are then processed through a grouped downsampling convolution to extract multi-scale features. This operation reduces the spatial resolution while effectively controlling the number of parameters through the grouping strategy. The spatial attention weight map and the multi-scale features are then summed element-wise with weights to obtain spatially enhanced features. This step achieves adaptive fusion of multi-scale features based on attention weights, enhancing the feature response of key regions.
[0028] Spatial augmentation features are fed into the PConv module for channel-level refinement. This convolutional unit supports two forward propagation modes: segmentation and concatenation, and slice replication. In segmentation and concatenation mode, the input tensor is precisely divided into 1 / 4 and 3 / 4 channels. Convolution, batch normalization, and ReLU activation are performed only on the 1 / 4 channel portion, while the 3 / 4 channel portion remains unchanged. Finally, the two parts are concatenated. In slice replication mode, the 1 / 4 channel portion of the input tensor is directly modified through tensor slicing operations, while the remaining channels remain unchanged. These operations ensure that the spatial augmentation features are refined to the desired channel dimensions. While some channels undergo deep processing, another portion of the channels retains rich semantic information from the LAE without loss. Finally, the concatenated features are fused with information and stabilized during training through a projection layer (1×1 convolution) and residual connections. The 1×1 convolution is responsible for mixing channel information from the two branches, while the conditional residual connections (enabled only when use_residual=True) provide direct gradient paths, effectively mitigating network degradation. The final output is a feature map that is optimized in both spatial and channel dimensions, possessing both strong representational capabilities and high computational efficiency.
[0029] In the task of small target detection in complex coal mine scenarios, as the input image passes through multiple layers of convolution and downsampling operations of convolutional neural networks, the feature map size gradually decreases, leading to the weakening or loss of local information of small targets. At the same time, there are still a large number of redundant or invalid features in the channels. Therefore, the LAPConv module proposed in this invention, which integrates the LAE module and the PConv module, achieves synergistic optimization in spatial feature representation and channel feature processing. It not only enhances the model's representation ability in complex environments and small target scenarios, but also reduces the number of model parameters and computational complexity, thus balancing the requirements of detection accuracy and lightweight design. This provides theoretical support for the reliable identification of small targets such as safety helmets and self-rescue devices in complex coal mine scenarios.
[0030] By combining the spatial adaptive enhancement of the LAE module with the selective channel extraction of PConv, the LAPConv module achieves collaborative optimization of small target features in complex coal mine scenarios, providing key technical support for reliable detection in resource-constrained environments. Through multi-scale feature extraction and batch normalization to enhance stability, and by combining spatial attention weights to weight input features, the model can better focus on important regions and improve the accuracy of small target detection.
[0031] PConv divides the input channels into convolutional branches and identity mapping branches in a 1:3 ratio (1 / 4 channels). This partially optimized strategy significantly reduces the computational burden from the source, allowing LAE to process rich features through spatial adaptive weighting. The input channels are divided into two parts through a channel segmentation strategy: 1 / 4 channels flow through computationally intensive 3×3 convolutions for depth processing, while the remaining 3 / 4 channels directly retain the original information enhanced by LAE through identity mapping. Finally, the advantages of both are combined by channel concatenation. This collaborative mechanism inherits the strong expressive power of LAE in spatial modeling and significantly reduces the overall computational complexity through PConv, ultimately achieving a significant improvement in computational efficiency while maintaining performance advantages.
[0032] In one embodiment, the MBPA detection head includes a regression branch and a classification branch.
[0033] like Figure 4 As shown, the regression branch includes a first MBPA_Block module, a second MBPA_Block module, and a first 1×1 convolutional layer connected in sequence.
[0034] The first MBPA_Block module has its expansion coefficient set to 1 and enables an efficient fusion attention mechanism. Internally, it consists of a 3×3 deep convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 1×1 partial convolutional layer, and a batch normalization layer. Feature calibration is achieved through a dual path of channel attention and spatial attention.
[0035] The expansion coefficient of the second MBPA_Block module is set to 2 and an efficient fusion attention mechanism is enabled. Its internal structure consists of a 1×1 partial convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 3×3 deep convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 1×1 partial convolutional layer, and a batch normalization layer. Feature calibration is achieved through a dual path of channel attention and spatial attention.
[0036] The first 1×1 convolutional layer is used to map the number of channels to four times the maximum value of the regression distribution, and outputs the distribution prediction of the bounding box.
[0037] The classification branch consists of a third MBPA_Block module, a 3×3 standard convolutional layer, and a second 1×1 convolutional layer connected in sequence.
[0038] The expansion coefficient of the third MBPA_Block module is set to 1 and the efficient fusion attention mechanism is disabled. Its internal structure consists of a 3×3 deep convolutional layer, a 1×1 partial convolutional layer, and a batch normalization layer.
[0039] The second 1×1 convolutional layer is used to compress the number of channels to the number of classes, and outputs the class prediction.
[0040] The MBPA detection head also includes: The splicing unit is used to splice the outputs of the regression and classification branches along the channel dimension.
[0041] The inference processing unit, in inference mode, is used to flatten the splicing result and split it into a bounding box distribution part and a category score part.
[0042] In this process, the bounding box distribution part is converted from discrete distribution to weighted coordinates through a distribution focus loss layer, and then the final bounding box coordinates are generated by combining the anchor point and the feature map stride through the distribution-to-bounding box function; the class score part is mapped to class probability through the Sigmoid activation function.
[0043] To address the issues of high computational complexity, large number of parameters, and limited feature representation capabilities of YOLOv11s in edge computing scenarios, this invention proposes a lightweight detection head architecture based on multi-branch parallel attention (MBPA), which significantly reduces computational overhead and model size while ensuring detection accuracy.
[0044] Starting from the two key branches of regression and classification paths, the detection head structure is reconstructed in a targeted manner. In the regression path of the object detection task, the MBPA module adopts a cascaded operation of extended convolution to increase feature dimension, depthwise convolution to extract spatial information, and compressed convolution to reduce the number of channels. The two MBPA modules are connected to the regression output convolution to form a complete regression path, which significantly reduces the computational cost while ensuring effective gradient propagation. In the classification path, a three-level cascaded structure of MBPA module, standard convolution and classification output convolution is adopted. This design achieves a significant reduction in computational load while ensuring classification function. The parameter scale is controlled by the channel optimization formula. By optimizing the regression path and classification path, the number of model parameters is significantly reduced without affecting the detection accuracy.
[0045] Traditional detectors typically employ a multi-layer standard convolutional design isomorphic to the regression path for classification paths. This homogenization fails to fully consider the fundamental differences between classification and regression tasks. The fully connected nature of standard convolutions in the channel dimension leads to a quadratic increase in the number of parameters with the number of channels. Furthermore, classification tasks are far less sensitive to spatial location than regression tasks. This computationally intensive design results in significant resource waste in edge computing scenarios. This invention specifically reconstructs the classification path structure of the detector head and proposes a hybrid cascaded design. Through modular combination, computational resources are allocated on demand, significantly improving computational efficiency without significantly reducing classification accuracy.
[0046] At the feature extraction level, depthwise separable convolution is used to replace the first-layer standard convolution. Depthwise separable convolution reduces computational complexity by decoupling spatial filtering and channel fusion. This design makes full use of the spatial invariance of classification features, and significantly reduces model complexity while maintaining effective feature extraction capabilities.
[0047] At the architecture composition level, a hybrid cascaded structure of depthwise separable convolution and standard convolution is constructed. To address the limitations of depthwise convolution in feature representation capabilities, a standard 3×3 convolutional layer is introduced in the second stage. Its dense spatial computation mode enhances the ability to capture texture details. This combination design retains the computational efficiency advantage of depthwise separable convolution while ensuring the full extraction of discriminative features through standard convolution, achieving a balanced distribution of computational load while ensuring feature quality.
[0048] At the channel configuration level, a channel optimization strategy based on task requirements is proposed. In traditional schemes, the classification path and regression path use the same channel configuration, which results in significant resource waste. By analyzing the feature redundancy characteristics of classification tasks, the number of intermediate feature channels is compressed, and the parameter scale is controlled from the source while maintaining the integrity of the classification decision boundary.
[0049] The process for detecting small targets in complex coal mine environments according to this invention is as follows: The input image is first fed into a feature extraction backbone network based on C3k2_MDConv. This network, based on the C3k2 structure of YOLOv11s, introduces MDCN to replace the standard convolution. MDCN dynamically adjusts the sampling position of the convolution kernel through learnable spatial offsets, and introduces learnable feature modulation weights for each sampling point. After nonlinear mapping by the Sigmoid function, the contribution of each sampling point is dynamically weighted. This design enables the convolution kernel to adaptively adapt to small targets such as well helmets and self-rescue devices according to the input features without significantly increasing computational complexity. At the same time, it effectively suppresses the interference of invalid features introduced when the sampling point shifts to the background area. It significantly improves the ability to extract detailed features such as edges and textures of small targets and the deformation adaptability from the source of the backbone network.
[0050] The multi-scale feature maps output by the backbone network are fed into the neck network. Utilizing the original FPN and PAN bidirectional pyramid architecture of YOLOv11, this achieves full interaction and deep fusion of high-level semantic information and low-level detailed features. Before the feature maps enter the detection head, this invention uses the LAPConv module to perform collaborative optimization of the spatial and channel dimensions of the features. This module first dynamically generates attention weights in the spatial dimension through LAE, adaptively weighting and fusing local regions of the feature maps, enabling the model to dynamically strengthen the representation of key regions based on the input content. Subsequently, PConv is introduced, dividing the input channels into convolutional branches and identity mapping branches in a 1:3 ratio. Only 1 / 4 of the channels undergoes 3×3 depth convolution processing, while the remaining 3 / 4 channels directly retain the original information enhanced by LAE through identity mapping. Finally, information fusion is achieved through channel concatenation and projection layers. This design, while maintaining low computational overhead, achieves an integrated collaborative design of feature enhancement and feature purification, significantly reducing redundant computation in the channel dimension from the computational source and significantly improving the model's feature discrimination ability for small targets in complex downhole backgrounds.
[0051] The multi-scale feature maps enhanced by the LAP module are fed into the MBPA detection head, which is heterogeneously designed to address the fundamental differences between classification and regression tasks. The regression path cascades two MBPA modules, each of which employs a cascaded operation of extended convolution to increase feature dimensions, depthwise convolution to extract spatial information, and compressed convolution to reduce the number of channels. Feature recalibration is achieved through parallel computation of channel attention and spatial attention paths, thereby enabling refined extraction of features in boundary regions. The classification path adopts a lightweight structure that combines depthwise separable convolution with standard convolution. Based on task requirements, the number of intermediate feature channels is compressed to 1 / 4 of the input channels. While maintaining the integrity of the classification decision boundary, the parameter scale is controlled from the source. This detection head architecture achieves an effective balance between high-precision boundary localization and efficient classification inference through the heterogeneous collaborative design of the regression and classification paths without adding any new detection head layers.
[0052] The optimization objective of the entire model is guided by a joint loss function, which combines bounding box regression loss and classification loss to jointly constrain the model and improve the overall detection performance in complex downhole environments.
[0053] This invention uses the YOLOv11s target detection framework as a benchmark and improves the model in three aspects: lightweight convolution, feature enhancement, and detection head, making it suitable for real-time detection of small targets in complex coal mine scenarios. The core inventive points of this invention include: (1) To address the problem of easy loss of details of small targets in complex coal mine scenarios, a C3k2_MDConv structure integrating MDCN was designed. By dynamically adjusting the sampling position of the convolution kernel and the feature modulation weight, the model’s adaptability to geometric deformation and detail perception of small targets is improved.
[0054] (2) In order to suppress feature redundancy and enhance key local features, a high-efficiency LAPConv feature enhancement mechanism that integrates LAE and PConv is proposed. While maintaining low computational overhead, it improves the model’s ability to distinguish features of small targets in complex downhole backgrounds.
[0055] (3) To address the problem of balancing computational efficiency and accuracy of the detection head, an MBPA detection head was designed. The regression path achieves refined boundary regression through MBPA module cascading, while the classification path adopts a lightweight cascaded convolutional structure to collaboratively optimize detection accuracy and inference speed.
[0056] In summary, the ML-YOLO small target detection algorithm for coal mines proposed in this invention, by introducing MDCN, enables the model to have a dynamic receptive field, which can adaptively fit the irregular edges and subtle textures of small targets such as safety helmets and self-rescue devices, significantly enhancing the ability to discriminate details of small-scale targets (less than 32×32 pixels). Through the collaboration of LAE and PConv, key areas are focused spatially, and redundancy is reduced in channels, effectively suppressing the interference of complex underground backgrounds and improving the model's feature expression capabilities. The MBPA dual-path heterogeneous detection head decouples regression and classification tasks, adopts parallel attention and a lightweight cascaded structure, optimizes the computation path, and greatly improves the inference efficiency of the detection head. The overall model maintains high accuracy while meeting the requirements of real-time processing speed, achieving significant advantages in four dimensions: accuracy, speed, lightweight, and robustness. It provides a high-performance, high-precision complete solution for solving the problem of real-time detection of small targets in complex coal mine scenarios, and has important application value and broad deployment prospects in intelligent mine construction and safety production monitoring systems.
[0057] The invention will now be described in detail with reference to simulation experiments.
[0058] (1) Technical effect verification data: The present invention has been fully experimentally verified on the self-built coal mine underground safety helmet and self-rescue device detection dataset. The key quantitative effect comparison is as follows.
[0059] (2) Module ablation experiment: The ablation experiment was conducted on the YOLOv11s model. The experimental results are shown in Table 1. In the table, A is the C3k2_MDConv module, B is the LAPConv module, C is the MBPA detection head, R is the recall rate, mAP50 is the average precision when the threshold is 0.5, GFLOPs is the floating-point computation cost, model size is the model size, and Parameters is the number of model parameters. As can be seen from the table, the three core innovative modules proposed in this invention have made independent and synergistic contributions to performance improvement. When the three modules work together, the model achieves optimal performance.
[0060] Table 1 Ablation Experiment Results (3) Comparison with the baseline model: YOLOv11s was selected as the baseline model. The comparison results are shown in Table 2. In the table, P is the average accuracy. It can be seen from the table that the model of the present invention (ML-YOLO) has achieved significant improvements in key evaluation indicators compared with the baseline model (YOLOv11s): mAP50 improved by 2.2%, P improved by 2.9%, and R improved by 2.4%. This directly proves the effectiveness of the technical solution of the present invention in the small target detection task in the complex scene of coal mine.
[0061] (4) To further verify the effectiveness of the ML-YOLO algorithm in small object detection tasks in complex coal mine scenarios, this invention (ML-YOLO) is compared with current mainstream lightweight object detection models (YOLOv5s, YOLOv8n, YOLOv10s, YOLOv11n, YOLOv11s, YOLOv12n) and Reference 1 (ZOU C, YU SQ, YU YK, et al. Side-ScanSonar Small Objects Detection Based on Improved YOLOv11[J]. Journal of Marine Science and Engineering, 2025, 13(1): A systematic comparative experiment was conducted using the method described in 162.), and the experimental results are shown in Table 2. As can be seen from Table 2, ML-YOLO achieved the best performance in both R and mAP50 key indicators. Specifically, P reached 91.3%, which is 5.6%, 5.8%, and 3.9% higher than YOLOv5s, YOLOv8n, and YOLOv11n, respectively; R improved to 76.4%, which is significantly better than YOLOv10s (69.8%) and YOLOv11n (69.6%); and mAP50 reached 85.1%, which is the best among all the compared methods, surpassing the current mainstream lightweight models and existing literature algorithms. Moreover, while maintaining high accuracy, the number of model parameters remained at 5.7M, indicating that this method achieved a significant improvement in accuracy while ensuring real-time performance.
[0062] Table 2 Comparison of detection results of different models Figure 5 This is a comparison chart of the detection visualization results of the present invention and different detection models in the prior art. In the chart, helmet is a safety helmet, self rescuer is a self-rescuer, and the numbers represent confidence levels. Scenario 1, Scenario 2 and Scenario 3 are all located in the mining face area, which has typical environmental interferences such as uneven brightness distribution, strong top lighting, and complex metal structures. Scenario 4 and Scenario 5 are located in the mine transport roadway area. The overall lighting in this scenario is extremely low, relying only on the weak light source on the top of the roadway or the equipment itself, resulting in low contrast and making it easy for the model to miss detections or reduce confidence levels.
[0063] As shown in the figure, in scenario one, the YOLOv8n model only detected the safety helmet but not the self-rescue device, resulting in a missed detection. The YOLOv10n and YOLOv11n models both identified highly reflective components such as pipe valves and metal connectors as self-rescue devices but failed to detect the safety helmet, resulting in false positives and false negatives. The YOLOv11s model detected both the safety helmet and the self-rescue device, but the confidence level for the self-rescue device was lower than that of the method of this invention. The method of this invention accurately detected both the safety helmet and the self-rescue device, and the confidence level was higher than that of other existing models.
[0064] In scenario two, the YOLOv8n model only detected the helmet but not the self-rescue device, resulting in a missed detection; the YOLOv10n model detected both the helmet and the self-rescue device, but its confidence level was lower than that of the method of this invention; the YOLOv11n model failed to detect either the helmet or the self-rescue device, resulting in a missed detection; the YOLOv11s model only detected the self-rescue device but not the helmet; the method of this invention accurately detected both the helmet and the self-rescue device, and its confidence level was higher than that of other existing models.
[0065] In scenario three, the YOLOv8n model detected both helmets and self-rescue devices, but its confidence scores were lower than those of the method of this invention; the YOLOv10n model only detected helmets; the YOLOv11n model detected both helmets and self-rescue devices, but its confidence scores were lower than those of the method of this invention; both the YOLOv11s model and the method of this invention detected helmets and self-rescue devices, and their confidence scores were the same, both higher than those of other models.
[0066] In scenario four, the YOLOv8n model detected two helmets and one self-rescue device, but its confidence level was lower than that of the method of this invention; the YOLOv10n model detected one helmet and one self-rescue device, but failed to detect all helmets, resulting in missed detections; the YOLOv11n model detected two helmets but failed to detect the self-rescue device, resulting in missed detections; the YOLOv11s model detected two helmets and one self-rescue device, but its confidence level was lower than that of the method of this invention; the method of this invention accurately detected two helmets and one self-rescue device, and its confidence level was higher than that of other existing models.
[0067] In scenario five, the YOLOv8n, YOLOv10n, YOLOv11n, YOLOv11s models and the method of this invention all detected two safety helmets and two self-rescue devices, but their average confidence scores were all lower than those of the method of this invention.
[0068] Therefore, the method of the present invention effectively suppresses the false detection problem caused by strong light reflection under strong light illumination, successfully distinguishes metal parts from self-rescue devices, and can accurately detect key targets such as safety helmets and self-rescue devices. The detection results are more stable and have higher confidence. Under extremely low light conditions, it effectively reduces the false judgment problem caused by shadows and background noise, demonstrating stronger low-light robustness and complex background separation ability.
[0069] (5) Experimental environment of the present invention: Ubuntu 22.04 operating system, PyTorch2.5.1 deep learning framework and Python 3.12 were used, and CUDA 12.4 was used for model training and testing. For details of the experimental environment and training parameters, please refer to Table 3 and Table 4.
[0070] (6) Dataset used in this experiment: The dataset used in the experiment was collected by deploying high-definition cameras in multiple underground coal mine scenes, including coal mining faces, transportation roadways, underground platforms, etc., to ensure that diverse scenes and activities could be captured.
[0071] Table 3 Experimental Environment Configuration Table 4 Experimental Model Setting Parameters
Claims
1. A lightweight small target detection method for complex coal mine scenarios, characterized in that, Includes the following steps: The image to be detected is input into the trained ML-YOLO model; Obtain and output the target detection results of the ML-YOLO model in the image to be detected; The ML-YOLO model is an improvement upon the YOLOv11s model. The ML-YOLO model includes a backbone network, a neck network, and a detection head. The improvements include: In the backbone network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module, and the convolutional module in the deep extraction stage of the YOLOv11s model is replaced with the LAPConv module. In the neck network, the C3k2 module in the YOLOv11s model is replaced with the C3k2_MDConv module; In the detection head, the original detection head of the YOLOv11s model is replaced with the MBPA detection head.
2. The lightweight small target detection method for complex coal mine scenarios according to claim 1, characterized in that, The C3k2_MDConv module is an improvement based on the C3k2 module, and the improvements include: Replace the Bottleneck unit in the C3k2 module with the Bottleneck_MDConv unit; The Bottleneck_MDConv unit is obtained by replacing the standard convolution in the Bottleneck unit with MDCN.
3. The lightweight small target detection method for complex coal mine scenarios according to claim 1, characterized in that, The feature processing in the LAPConv module includes: The input feature map is adaptively weighted and fused in the spatial dimension by using lightweight adaptive attention, and then the input feature map after adaptive weighting and fusion is purified in the channel dimension by partial channel convolution to obtain the enhanced feature map.
4. The lightweight small target detection method for complex coal mine scenarios according to claim 3, characterized in that, The LAPConv module includes an LAE module and a PConv module connected in sequence; The feature processing in the LAE module includes: The input feature map is subjected to average pooling, and contextual information is extracted through 1×1 convolution to obtain intermediate features; The intermediate features are reorganized from the spatial dimension to the block dimension through a recombination operation; The recombined features are normalized using the Softmax function to generate a spatial attention weight map. The input feature map is subjected to grouped downsampling convolution to extract multi-scale features; The spatial attention weight map and the multi-scale features are summed element-wise to obtain the spatial enhancement features.
5. A lightweight small target detection method for complex coal mine scenarios according to claim 4, characterized in that, The feature processing in the PConv module includes: The spatial enhancement features are divided into convolutional branches and identity mapping branches according to channels; Perform a convolution transformation on the convolution branch; The transformed convolutional branch and the identity mapping branch are concatenated and fused through channels to obtain intermediate fused features; The intermediate fused features are mixed with channel information through a projection layer, and the final enhanced feature map is output by combining residual connections.
6. The lightweight small target detection method for complex coal mine scenarios according to claim 5, characterized in that, Performing a convolution transformation on the convolution branch includes: The convolutional branch is subjected to 3×3 depthwise convolution, batch normalization, and ReLU activation.
7. The lightweight small target detection method for complex coal mine scenarios according to claim 1, characterized in that, The MBPA detection head includes a regression branch and a classification branch; The regression branch includes a first MBPA_Block module, a second MBPA_Block module, and a first 1×1 convolutional layer connected in sequence; The expansion coefficient of the first MBPA_Block module is set to 1 and an efficient fusion attention mechanism is enabled. Its internal structure consists of a 3×3 deep convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 1×1 partial convolutional layer, and a batch normalization layer. Feature calibration is achieved through a dual path of channel attention and spatial attention. The expansion coefficient of the second MBPA_Block module is set to 2 and an efficient fusion attention mechanism is enabled. Its internal structure consists of a 1×1 partial convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 3×3 depth convolutional layer, a batch normalization layer, a ReLU6 activation layer, a 1×1 partial convolutional layer, and a batch normalization layer. Feature calibration is achieved through a dual path of channel attention and spatial attention. The first 1×1 convolutional layer is used to map the number of channels to four times the maximum value of the regression distribution, and output the distribution prediction of the bounding box.
8. A lightweight small target detection method for complex coal mine scenarios according to claim 7, characterized in that, The classification branch includes a third MBPA_Block module, a 3×3 standard convolutional layer, and a second 1×1 convolutional layer connected in sequence; The expansion coefficient of the third MBPA_Block module is set to 1 and the efficient fusion attention mechanism is disabled. Its internal structure consists of a 3×3 depth convolutional layer, a 1×1 partial convolutional layer, and a batch normalization layer. The second 1×1 convolutional layer is used to compress the number of channels to the number of categories, and outputs the category prediction.
9. A lightweight small target detection method for complex coal mine scenarios according to claim 1, characterized in that, The MBPA detection head also includes: A splicing unit is used to splice the outputs of the regression branch and the classification branch along the channel dimension; The inference processing unit, in inference mode, is used to flatten the splicing result and split it into a bounding box distribution part and a category score part; The bounding box distribution part sequentially passes through a distribution focus loss layer to convert the discrete distribution into weighted coordinates, and then the distribution to bounding box function combines the anchor point and the feature map stride to generate the final bounding box coordinates. The category score portion is mapped to category probability using the Sigmoid activation function.
10. A lightweight small target detection device for complex coal mine scenarios, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-9.