Mine small target detection method based on deformable convolution and residual structure

Abstract

The invention discloses a mine small target detection method based on deformable convolution and a residual structure, belongs to the technical field of underground small target detection, and further improves the detection precision and robustness of a small target by introducing multi-scale feature fusion and an attention mechanism. The method comprises the following steps: a backbone network reinforces cooperative perception of channel, space and position information in a feature extraction stage by fusing an MLCA attention mechanism, and effectively inhibits background noise interference; on the basis of a DPC-Block multi-scale feature fusion network, the relevance between low-order detail information and high-order semantic features is reserved through cross-layer feature interaction; the detection layer adopts a ShapeIoU loss function to optimize bounding box regression precision and accelerate model convergence; according to the invention, high target detection precision can be realized, and the problem of missing detection of small targets is effectively improved; and an effective solution is provided for small target detection in a complex scene in collaborative optimization of an attention mechanism and feature fusion and a geometric perception loss function.

Description

Technical Field

[0001] The present invention belongs to the technical field of underground small target detection, and in particular relates to a method for detecting small targets in mines based on deformable convolution and residual structure. Background Art

[0002] With the advent of the era of smart cities and smart industries, intelligent object detection algorithms have been gradually applied in many industries, including industrial inspection, behavior recognition, fire safety, traffic management, and biomedicine. As application scenarios continue to expand, the performance requirements for object detection models continue to increase, especially in the detection of small objects, which still faces significant challenges.

[0003] Currently, algorithms used for object detection fall into two main categories: single-stage and two-stage. Single-stage object detection algorithms generally offer advantages such as speed, ease of deployment, and high real-time performance. Classic single-stage object detection algorithms include the YOLO (You Only Look Once) family of algorithms and the SSD (Single Shot Detector) family of algorithms. Two-stage object detection algorithms, consisting of two stages: candidate bounding box generation and classification refinement, achieve higher accuracy in object localization and classification and are more adaptable to handling multi-scale objects. However, two-stage approaches often suffer from high complexity, difficulty in deployment, and unsuitability for high-performance real-time applications. Classic two-stage object detection algorithms include the R-CNN (Region with CNN feature) family of algorithms and the Faster R-CNN algorithm. Given the current demands of industrial scenarios, which place higher demands on real-time object detection and rapid deployment, single-stage object detection algorithms offer advantages over two-stage algorithms. In recent years, with the continuous development of the YOLO network, many researchers have proposed target detection models based on the YOLO network for application in different scenarios. However, the target detection methods for personnel and equipment in coal mines still face the challenge of unsatisfactory algorithm false detection and missed detection rates. In addition, due to the complex environment and poor lighting conditions of many small target detections, the model still faces significant challenges in real-time performance and high detection accuracy.

[0004] In summary, although the current single-stage small target detection model has made considerable research progress, further research is still needed on the following two issues: (1) How to effectively optimize the feature extraction network and feature fusion network in the single-stage model to adaptively mine the detailed features of smaller targets while effectively suppressing the interference of background noise to achieve high detection accuracy; (2) How to improve the robustness of the small target detection model in complex environments to ensure that the model can efficiently and accurately detect small targets in complex scenes such as mines, especially high-contrast small targets such as miners' helmets.

[0005] To address these problems, a small target detection structure YOLOv8-DPMS is proposed. By introducing multi-scale feature fusion and attention mechanism, the detection accuracy and robustness of small targets are further improved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a mine small target detection method based on deformable convolution and residual structure to address the problems of missed detection, false detection and low detection accuracy of current small target detection algorithms for coal mine underground data sets.

[0007] The present invention adopts the following technical solutions to solve the above technical problems:

[0008] The method for detecting small targets in mines based on deformable convolution and residual structure specifically includes:

[0009] (1) Image preprocessing module: preprocess the collected original target image;

[0010] (2) Small target feature extraction network module based on MLCA attention mechanism: This module focuses on channel, spatial, and position information during the small target feature extraction process, suppresses background noise interference, and enhances the expressiveness of the extracted features. The preprocessed image is input into this module to suppress the interference of background noise in the small target image and extract features with high representational capabilities. The features at different layers are then input into the feature enhancement fusion module.

[0011] (3) Small target feature fusion and reinforcement network module based on deformable convolution and residual structure: The different scale features output by the C2f module and the spatial pyramid pooling module in the feature extraction network are input into the feature fusion and reinforcement network, so that the shallow and deep features in the feature extraction network are fused and reinforced, while learning high-order semantic features, the rich detail information represented by the low-order features is retained, thereby effectively improving the accuracy of small target detection;

[0012] (4) Target classification and bounding box prediction network: Based on the output of the four MLCA modules of the feature fusion enhancement network module, the prediction network can focus on important features, forming four target prediction branches, and using different anchor boxes to predict targets of different sizes. Each anchor box corresponds to a target of a specific size. During the training process, in order to speed up the convergence of the target detection model and improve the regression accuracy of the target prediction box, the ShapeIOU loss function is used to optimize the original loss function of the network to improve the detection accuracy.

[0013] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0014] The present invention provides an underground small target detection method based on deformable convolution and hybrid local channel attention, which further improves the detection accuracy and robustness of small targets by introducing multi-scale feature fusion and attention mechanism; specifically as follows: a backbone network based on the MLCA attention mechanism, which is constructed based on two-dimensional convolution, C2f module, spatial pyramid pooling module, and MLCA attention mechanism module, to achieve the integration of attention to channel, space and position information in the process of small target feature extraction, suppress the interference of background noise and enhance the expressive ability of the extracted features; a small target feature fusion and enhancement module DPC-Block based on deformable convolution and residual structure is proposed, so that the shallow and deep features in the feature extraction network are fused and enhanced, while learning high-order semantic features, the rich detail information represented by the low-order features is retained, thereby improving the detection accuracy; target classification and bounding box prediction module; in order to accelerate the convergence speed of the target detection model and improve the regression accuracy of the target prediction box, the ShapeIoU loss function is used so that YOLOv8-DPMS can better classify and locate the target from the fused feature map obtained from the feature fusion module. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the positions of standard convolution and deformable convolution of the present invention;

[0016] Figure 2 3×3 deformable convolution diagram of the present invention;

[0017] Figure 3 This is a comparison between the Bottleneck module of the present invention and the DPC-Block module;

[0018] Figure 4 It is a schematic diagram of the hybrid local channel attention (MLCA) algorithm of the present invention;

[0019] Figure 5 This is the ShapeIOU loss structure diagram of the present invention;

[0020] Figure 6 It is the YOLOv8-DPMS algorithm structure of the present invention;

[0021] Figure 7 This is the internal structure diagram of the Conv module of the present invention;

[0022] Figure 8 This is a diagram of the internal structure of the C2f module of the present invention;

[0023] Figure 9 This is a diagram showing the internal structure of the MLCA module of the present invention;

[0024] Figure 10 It is the internal structure diagram of the SPPF module of the present invention;

[0025] Figure 11 It is a small target feature extraction network based on the MLCA attention mechanism of the present invention;

[0026] Figure 12 It is a structural diagram of the CDPC module of the present invention;

[0027] Figure 13 The present invention is based on the feature fusion and enhanced network structure of DPC-Block;

[0028] Figure 14 This is a diagram showing the internal structure of the improved detection head of the present invention;

[0029] Figure 15 It is the detection layer network structure of the present invention;

[0030] Figure 16 It is a schematic diagram of part of the Coal-H data set of the present invention;

[0031] Figure 17 It is the map@0.5 result diagram of the ablation experiment of each category of the present invention. DETAILED DESCRIPTION

[0032] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings:

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The present invention is described in detail below based on the drawings and preferred embodiments. The purpose and effect of the present invention will become more clear. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0034] This paper proposes a small object detection architecture, YOLOv8-DPMS, which further improves the detection accuracy and robustness of small objects by introducing multi-scale feature fusion and attention mechanisms. YOLOv8-DPMS is a YOLOv8-based object detection model that incorporates the DPMS (Dynamic Position-aware Multi-scale) mechanism to improve the accuracy and efficiency of object detection. The main contributions of this method are as follows:

[0035] (1) A backbone network based on the MLCA attention mechanism, which is constructed based on two-dimensional convolution, C2f module, spatial pyramid pooling module, and MLCA attention mechanism module, realizes the integration of attention to channel, space and position information in the process of small target feature extraction, suppresses the interference of background noise and enhances the expressiveness of the extracted features.

[0036] (2) A small target feature fusion and enhancement module DPC-Block based on deformable convolution and residual structure is proposed, which enables the fusion and enhancement of shallow and deep features in the feature extraction network, while retaining the rich detail information represented by low-level features while learning high-level semantic features, thereby improving detection accuracy;

[0037] (3) Target classification and bounding box prediction module. To accelerate the convergence of the target detection model and improve the regression accuracy of the target prediction box, the ShapeIoU loss function is used so that YOLOv8-DPMS can better classify and locate targets from the fused feature map obtained from the feature fusion module.

[0038] The rest of the present invention is organized as follows: the relevant principles of DPC-Block, MLCA attention mechanism module, and ShapeIOU are introduced in detail, with an emphasis on the proposed YOLOv8-DPMS small target detection algorithm. Target detection experiments are conducted on the original dataset, and the experimental results are analyzed.

[0039] Theoretical Background

[0040] DPC-Block: The Bottleneck residual module in feature fusion networks is primarily used to address problems such as vanishing gradients, feature information loss, and high computational complexity in deep networks. In deep learning models, as the number of network layers increases, gradients can gradually vanish or explode during backward propagation, making training difficult. The Bottleneck architecture uses residual connections to allow gradients to propagate directly from the later layers to the earlier ones, effectively alleviating the vanishing gradient problem. These skip connections allow the network to learn a residual mapping during training, rather than directly learning a complete input-to-output mapping, making the optimization process more efficient and improving training stability. The feature maps of the Bottleneck residual module are typically processed through a series of convolutions, normalization, and activation operations. Each Bottleneck residual module contains two 3x3 convolutional layers, which transform the input feature maps to extract higher-level feature representations. However, this can reduce model accuracy for objects with irregular scale variations or large scale differences.

[0041] Deformable convolution is a technique that extends traditional convolution operations, aiming to enhance the performance of convolutional neural networks (CNNs) in processing complex shapes and spatial deformations. Traditional convolution operations use fixed-size convolution kernels when processing images, and the convolution kernels perform the same operation at each position. However, the targets or features in the image may have different shapes, scales, or positions, and a fixed convolution window may not be able to effectively capture these deformed features. Deformable convolution adds a 2D offset to the regular grid sampling positions in the standard convolution. It allows the sampling grid to deform freely and dynamically adjusts the sampling position of the convolution kernel by learning a set of offsets. Unlike the traditional fixed sampling method of convolution kernels, deformable convolution allows the sampling position of the convolution operation to be offset, such as Figure 1 shown. Figure 1 Schematic diagram of the positions of standard convolution and deformable convolution

[0042] Deformable convolution is performed through an independent network branch, which performs a convolution operation on the input feature map to generate an offset feature map with 2N channels. The offset of the pixel points in the feature map is optimized and updated using a bilinear interpolation backpropagation algorithm. The offset is usually a floating point number. The input feature map is convolved with the updated offset to obtain the output feature map. This enhances the performance of convolutional neural networks (CNNs) in processing complex shapes and spatial deformations. The specific process is as follows: Figure 2 shown. Figure 2 Schematic diagram of 3×3 deformable convolution.

[0043] The eigenvalue output formula of deformable convolution kernel sampling is:

[0044]

[0045] Among them, x is the input feature map, w k represents the weight of k position, p k represents the pre-offset, p is the center position of the sampling point on the feature map, Δp k is the offset relative to the center position p. When it is 0, it is the standard convolution kernel. Δm k Represents the learnable weight, ranging from 0 to 1. For sampling points that do not need to be adjusted, their weights are set to 0, making the deformation of the convolution kernel more flexible.

[0046] Therefore, this paper uses deformable convolution to propose a feature fusion module DPC-Block that enhances multi-scale feature fusion and captures complex geometric deformations. The comparison between the common Bottleneck module and the DPC-Block module we proposed is as follows: Figure 3 shown.

[0047] DPC-Block employs a multi-level feature extraction and fusion strategy to enhance the model's ability to perceive objects with complex geometric deformations. First, a set of 3×3 deformable convolutional modules is introduced. Through a bias learning mechanism, the model's adaptability to non-rigid deformations is enhanced, enabling it to more effectively model the relationships between objects with complex geometric structures, thereby improving its ability to capture deformed objects. Next, a set of 1×1 basic convolutional modules is added to adjust low-level channel information, reducing computational complexity while maintaining effective feature representation.

[0048] Furthermore, to enhance the ability to fuse multi-scale information, the DPC-Block uses Concat to connect feature maps of different scales, fully integrating low-level local detail information with high-level semantic information. This enriches target details while minimizing spatial information loss, thereby improving accurate target positioning. This structural design not only effectively enhances the ability to fuse multi-scale features but also improves the model's accuracy in capturing complex geometrically deformed targets, further enhancing the performance of target detection tasks.

[0049] MLCA Attention Mechanism (Mixed Local Channel Attention):

[0050] In actual image target detection tasks, the same target will have different sizes and shapes when viewed at different camera positions, thereby reducing the accuracy of target detection. In addition, in real scenes, there are occlusion problems between targets to be detected and between targets to be detected and unrelated targets. Small targets face more such problems in real scenes. In this regard, introducing an attention mechanism into the network can significantly reduce the negative impact of the above situation on the detection results and further improve the performance of the model. In order to effectively model features at different levels, many researchers have proposed various variants of attention mechanisms, including channel attention and spatial attention. However, most channel attention mechanisms only contain channel feature information and ignore spatial feature information, resulting in poor model representation or target detection performance. Spatial attention modules are often complex and costly. Therefore, the present invention uses a lightweight mixed local channel attention mechanism MLCA (Mixed Local Channel Attention), which enhances the detection capability of small targets by combining local spatial information and channel features. Figure 4 Schematic diagram of the mixed local channel attention (MLCA) algorithm.

[0051] like Figure 4As shown in Figure 2, the MLCA attention mechanism works as follows. First, the input image is passed through the convolutional layer for feature extraction to obtain a feature map. In the figure, Conv1d represents a one-dimensional convolution operation, and the kernel size k depends on the channel dimension C. This setting means that in the process of capturing local cross-channel interactions, the focus is only on the relationship between each channel and its k neighboring channels. The choice of k is expressed by formula (2):

[0052]

[0053] Where C represents the number of channels, k represents the size of the convolution kernel, and γ and b are hyperparameters, both preset to 2. k must be an odd number. If k is an even value, it is adjusted in increments of 1.

[0054] For each feature map position, a local area is defined by setting a local receptive field. For each local area, its position features are calculated, including local center point and boundary features. The local center point feature represents the center point of the area, while the boundary feature represents the relative position of the area to the overall image boundary. For each local area, its channel features are calculated to capture information from different channels. Two parallel fully connected layers are used to process position features and channel features respectively. The correlation between the learned position features and channel features is obtained through the fully connected layer, and the position attention weight and channel attention weight are obtained. The position attention weight and the channel attention weight are multiplied to obtain the final mixed local channel attention weight. Then, this weight is multiplied with the feature map to fuse features from different channels and positions. Finally, the fused feature map is input into the subsequent network layer for further processing and classification tasks.

[0055] By hybridizing the local channel attention mechanism, the model can more accurately focus on the importance of different channels in the image and extract more discriminative features, thereby improving and enhancing the performance and effect of small object detection tasks.

[0056] ShapeIOU: Most of the current mainstream object detection algorithms use CIoU as the loss function. This loss function has accurate target positioning and considers the integrity of the target, but it has certain shortcomings for small and irregular targets. Shape-IoU is introduced in the YOLOv8-DPMS algorithm to consider the shape and scale of the boundary, which is more accurate. The Shape-IoU loss structure diagram is shown below. Figure 5 As shown; Figure 5 This is the ShapeIOU loss structure diagram.

[0057] ShapeIoU focuses on the shape and size information of the bounding box itself and integrates this information into the IoU loss function. The calculation formula is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Where scale is the scale factor, which is related to the scale of the target in the dataset. ww and hh are the weight coefficients in the horizontal and vertical directions respectively, and their value range is related to the shape of the GT box. The corresponding bounding box regression loss is as follows:

[0065] L Shape-IoU =1-IoU+distance shape +0.5×Ω shape (9)

[0066] By considering the outline of the target, ShapeIoU can effectively improve the detection accuracy of small targets and reduce the detection error caused by inaccurate framing.

[0067] Proposed YOLOv8-DPMS algorithm: YOLOv8-DPMS algorithm framework

[0068] Figure 6 This is the YOLOv8-DPMS algorithm structure. To address the problems of missed detection, false detection, and low detection accuracy in current small target detection algorithms for coal mine datasets, this paper studies feature extraction based on convolutional neural networks, embedding attention mechanisms, and multi-scale feature fusion, and proposes a small target detection algorithm, YOLOv8-DPMS. The algorithm consists of four modules:

[0069] (1) Image preprocessing module. This module preprocesses the collected original target images and uses the Mosaic data enhancement method to process the original images. The Mosaic method uses four images and splices them in a random scaling, cropping and arrangement manner to combine multiple images. This can expand the data set while improving the model training speed and reducing memory requirements.

[0070] (2) Small target feature extraction network module based on MLCA attention mechanism. Figure 6As shown in the figure, the network consists of five Conv layers, four C2f modules, four MLCA modules, and one Spatial Pyramid Pooling (SPPF) module. This module focuses on channel, spatial, and positional information during small object feature extraction, suppressing background noise interference and enhancing the expressiveness of the extracted features. The preprocessed image is fed into this module to suppress background noise interference in small object images and extract highly expressive features. Features from different layers are then fed into the Feature Enhancement Fusion module.

[0071] (3) A small target feature fusion and reinforcement network module based on deformable convolution and residual structure. The network consists of two upsampling modules, six CDPC modules, and three Conv modules. The different scale features output by the C2f module and SPPF module in the feature extraction network are input into the feature fusion and reinforcement network, so that the shallow and deep features in the feature extraction network are fused and reinforced. While learning high-order semantic features, the rich detail information represented by low-order features is retained, thereby effectively improving the accuracy of small target detection.

[0072] (4) Target classification and bounding box prediction network. This network is based on the output of the four MLCA modules of the feature fusion enhancement network module, which enables the prediction network to focus on important features and form four target prediction branches. Different anchor boxes are used to predict targets of different sizes. Each anchor box corresponds to a target of a specific size. During the training process, in order to accelerate the convergence of the target detection model and improve the regression accuracy of the target prediction box, the ShapeIOU loss function is used to optimize the original network loss function to improve the detection accuracy.

[0073] Feature extraction network based on MLCA attention mechanism:

[0074] Depend on Figure 6 As shown in Figure 2, the small target feature extraction network based on the MLCA attention mechanism consists of 5 Conv modules, 4 C2f modules, 4 MLCA modules, and 1 spatial pyramid pooling module (SPPF). The internal structures of the Conv module, C2f module, MLCA module, and spatial pyramid pooling module are shown in Figure 2. Figure 7-10 As shown in the figure. The Conv module is a commonly used basic module, which consists of two-dimensional convolution (Conv2d), batch normalization (BN) and activation function SiLU, which enhances the nonlinear ability of convolution operation. The most basic unit ensures that the model can capture low-level features. The internal structure of the C2f module is shown in the figure. Figure 8As shown in the figure, during the forward propagation process, C2f first divides the input feature map into two parts, and then performs convolution and Bottleneck processing on these two parts respectively. The processed feature maps are recombined and then generated through the final convolution layer to enhance the expressiveness of the feature maps and improve the model's ability to recognize complex targets. The internal structure of the MLCA module is shown in the figure. Figure 9 As shown in Figure 2, the module adopts a multi-scale channel-aware strategy, taking into account both channel information and spatial information, and combining local information with global information to improve the network's feature extraction capability and further enhance the network's recognition accuracy for small targets. The internal structure of the SPPF module is shown in Figure 2. Figure 10 As shown in the figure, this module receives the feature map of the previous layer of SPPF as input and uses multiple small pooling layers to replace the pooling operation of a single large core. The main advantage of this design is that through pooling operations at different scales, SPPF can extract richer multi-scale features and generate a fixed-length feature vector by splicing the outputs of each pooling layer, ultimately enhancing the network's adaptability to targets of different sizes, thereby further improving the effect of small target detection. Figure 7 This is the internal structure diagram of the Conv module. Figure 8 This is the internal structure diagram of the C2f module. Figure 9 This is the internal structure diagram of the MLCA module. Figure 10 This is the internal structure diagram of the SPPF module. Figure 11 It is a small target feature extraction network based on MLCA attention mechanism.

[0075] Based on the Conv module, C2f module, MLCA module and SPPF module, the present invention constructs a small target feature extraction network based on the MLCA attention mechanism, whose structure is as follows: Figure 11 By extracting features from images using this structure, we can integrate channel, spatial, and position information into the process of extracting small target features, better suppress the interference of background noise, and enhance the expressiveness of the extracted features, providing features with strong expressiveness for subsequent feature enhancement and fusion tasks.

[0076] Feature Fusion Enhancement Network Based on DPC-Block: Shallow neurons in convolutional neural networks can only learn low-level features of simple object details (such as edges, textures, etc.), but this characteristic shows significant limitations when processing targets in complex environments. For example, the target may be affected by interference such as occlusion, rotation, or deformation. Especially in small target detection tasks, traditional convolution kernels are unable to capture detailed features, which easily leads to the loss of important information. This loss of detail directly affects the accuracy of small target detection and makes it difficult for the model to maintain robustness in complex scenes. Deformable convolution effectively adapts to irregular deformations and complex backgrounds of the target through dynamic sampling capabilities, helping to capture more detailed feature information. Therefore, a feature fusion enhancement module (DPC-Block) is used to improve the accuracy of small target detection and feature expression capabilities.

[0077] The complete network structure is as follows Figure 13 As shown in Figure 2, the network consists of 6 CDPC modules, 3 upsampling modules, 3 Conv modules, and 6 Concat modules. The internal structure of the CDPC module is shown in Figure 2. Figure 12 As shown in the figure, through dynamic feature fusion and cross-scale interaction, the target detection capability in complex scenes is effectively enhanced, providing strong support for processing small targets. Figure 12 This is the CDPC module structure diagram; Figure 13 Strengthen the network structure for feature fusion based on DPC-Block.

[0078] Improvements to the detection layer and loss function: Figure 14 This is a diagram of the internal structure of the improved detection head; Figure 15 It is the network structure of the detection layer;

[0079] In order to improve the accuracy and robustness of small target detection, the detection head structure and loss function are improved and designed. At the same time, in order to make the detection head pay more attention to important features and thus detect small targets, the MLCA attention mechanism is added before detection. The detection layer network structure is as follows: Figure 15 As shown in Figure 2. The improved detection head structure includes multiple convolutional layers (Conv) and a two-dimensional convolutional layer (Conv2d), which is divided into a bounding box regression branch and a classification branch, as shown in Figure 2. Figure 14 As shown, the regression branch uses a combined loss based on ShapeIOU and Distribution Focal Loss (DFL) to optimize bounding box localization accuracy; the classification branch introduces Binary Cross Entropy Loss (BCE) to improve classification performance. ShapeIOU effectively solves the problem of vanishing gradients in small overlapping areas of traditional IoU, while DFL further refines the predicted distribution of the bounding box. This effectively improves the model's convergence time and provides strong support for processing small objects.

[0080] Experimental Analysis: Introduction to Experimental Dataset

[0081] To verify the superior performance of the proposed YOLOv8-DPMS algorithm in small target detection and effectively enhance its feature extraction and multi-scale feature fusion capabilities for small targets with lower model parameters, this study conducted experimental verification using the self-built Coal-Helmet dataset. The experiments evaluated the algorithm's small target detection performance in complex environments, specifically focusing on improving the model's sensitivity and accuracy for small targets under environmental constraints.

[0082] Coal-H self-made dataset: This dataset Coal-H is specially built for the task of detecting small targets in underground mines, based on the actual collected underground working environment data. The dataset contains 500 experimental data, and each image is annotated with the specific location and category of the target to ensure the accuracy and completeness of the dataset. The training set and test set are divided into 8:2 ratios. Some datasets are as follows: Figure 16 As shown, the categories include pipelines, tracks, miners, safety helmets, anti-impact clothing, belt areas, and no anti-impact clothing. Figure 16 Schematic diagram of part of the Coal-H dataset;

[0083] Experimental evaluation indicators:

[0084] The main evaluation indicators for target detection model performance are precision and recall.

[0085] The present invention uses three indicators, namely, precision, recall, and mean average precision (mAP), which are widely used in the research field, to measure the performance of the model target detection. Among them, precision and recall are often regarded as evaluation indicators for binary classification problems, with the class of interest as the positive sample and the class of no interest as the negative sample. In target detection, samples are often divided according to the value of intersection of union. That is, when the ratio of the intersection and union of the candidate box and the original marked box is greater than a certain set value, the candidate box is judged to be a positive sample, otherwise it is a negative sample. Therefore, the prediction results are divided into the following four categories: TP (True Positives), TN (True Negatives), FP (False Positives) and FN (False Negatives), as shown in Table 1.

[0086] Table 1

[0087]

[0088] Precision measures the similarity between the model's detection results and the true annotations, focusing on predicted positive samples and actual positive and negative samples. Its calculation formula is as follows. As can be seen, a higher precision value results in a lower FP, and the purity of the predicted positive samples becomes higher, meaning fewer false positives are detected.

[0089]

[0090] Recall refers to the ratio of correctly detected targets to the actual number of targets, focusing on the predicted positive and negative samples. Its calculation formula (11) is as follows. It can be seen that the larger the recall value, the smaller the FNFN, the fewer positive samples are predicted as negative samples, and the fewer missed detections.

[0091]

[0092] mAP represents the average precision of all categories in the dataset and is often used to reflect the accuracy of the entire model. mAP is the average AP value of each category, and its formulas are expressed in (12) to (14). The larger the mAP value, the larger the area enclosed by the PR curve and the coordinate axis. The commonly used mAP0.5 form represents the mAP value when the IOU threshold is 0.5, and the PR curve is obtained by using recall and precision as the horizontal and vertical coordinates respectively.

[0093]

[0094]

[0095]

[0096] (2) In terms of detection speed, the parameter quantity (Params) model evaluation index is used. The parameter quantity (Params) refers to the number of parameters contained in the model. The calculation formula is as follows:

[0097] Params=C in ×C out ×K×K (15)

[0098] In the above formula, K represents the size of the convolution kernel, C in and C out Respectively represent the number of input and output channels.

[0099] Experimental results and analysis:

[0100] To verify the effectiveness and advantages of the proposed YOLOv8-DPMS, comparative experiments with YOLOv5, YOLOv6, YOLOv7, YOLOv9, YOLOv10, and YOLOv11, as well as ablation experiments, were conducted on the Coal-H dataset. The models are shown in Table 2. Object detection experiments were conducted on the Coal-H dataset using the PyTorch deep learning framework and Python as the development language. The hardware and software environment configurations are shown in Table 3. The training strategy is as follows: The BatchSize parameter was 16, the initial learning rate was 0.01, the decay coefficient was 0.0005, and the minimum learning rate was 0.0005. The number of training epochs was set to 300.

[0101] Table 2

[0102]

[0103] Table 3

[0104]

[0105]

[0106] The comparative experimental results are shown in Table 4. The comparative experimental results for the small target category of helmets are shown in Table 5.

[0107] Table 4

[0108]

[0109] Table 5

[0110]

[0111] Tables 4 and 5 show the experimental results of seven comparison algorithms on the Coal-H dataset, showing four evaluation index data (mAP@0.5, mAP@0.5-0.95, Recall, Prams). According to the experimental data, we can conclude that:

[0112] Table 4 shows the experimental results for the overall category, demonstrating that YOLOv8-DPMS achieves significant improvements in metrics such as Recall, Map@0.5, and Map@0.5-0.95. Compared to YOLOv5, YOLOv8-DPMS improves Recall by 2.1% and Map@0.5 by 3.4%. In particular, YOLOv8-DPMS achieves 0.787 in Map@0.5-0.95, a measure of overall detection accuracy, surpassing both YOLOv7 (0.755) and YOLOv9 (0.779). This demonstrates that YOLOv8-DPMS achieves superior performance in fine-grained object detection, particularly in object localization and classification accuracy. This demonstrates that its innovative module design (such as DPC-Block, ShapeIoU, and MLCA) significantly improves detection capabilities without significantly increasing computational cost.

[0113] As shown in Table 5, YOLOv8-DPMS also shows a significant advantage in the detection of small targets such as helmets. First, the Recall index increased from 0.846 in YOLOv5 to 0.896, an increase of 5%. This change shows that YOLOv8-DPMS has a stronger detection capability for small targets and can more comprehensively identify small targets such as helmets. This is especially important for detection tasks in mine working environments with many occlusions and small-sized targets. Second, Map@0.5 increased from 0.834 in YOLOv6 to 0.942, an increase of 10.8%. This result shows that YOLOv8-DPMS has made significant improvements in the precise positioning of small targets, especially for targets such as helmets that usually have a strong contrast with the background, and the accuracy has been greatly improved. Finally, on Map@0.5-0.95, YOLOv8-DPMS achieved 0.556, which is significantly better than YOLOv5 (0.462) and YOLOv6 (0.432), indicating that its overall performance in processing small objects has also been greatly improved.

[0114] By comparison, YOLOv8-DPMS has more significant improvements in small target detection, especially in accuracy (Map@0.5) and comprehensive accuracy (Map@0.5-0.95), indicating that it has stronger adaptability to small targets in complex mine environments and can better cope with the diversity and complexity of small targets in mine operation scenarios.

[0115] The ablation experiment results are shown in Table 6. The ablation experiment results for the small target category helmet are shown in Table 7.

[0116] Table 6

[0117]

[0118] Table 7

[0119]

[0120] Tables 6 and 7 show the experimental results of the four comparison algorithms on the Coal-H dataset, showing four evaluation index data (mAP@0.5, mAP@0.5-0.95, Recall, Prams). The Map@0.5 index of each category is as follows: Figure 17 As shown, according to the experimental data, it can be concluded that:

[0121] The experimental results for the overall category in Table 6 show that object detection performance gradually improves from Model A to Model D. Specifically, Recall increases from 0.936 to 0.956, Map@0.5 improves from 0.963 to 0.981, and Map@0.5-0.95 improves from 0.762 to 0.787. In particular, Model D improves by 2.5% in the Map@0.5-0.95 metric compared to the base model, Model A, and by 1.8% in the Map@0.5 metric. This trend demonstrates that the introduction of modules such as DPC-Block, ShapeIoU, and MLCA significantly enhances feature extraction and multi-scale object representation capabilities. Furthermore, while the number of parameters increases from 3.0M to 4.2M, the performance gain far outweighs this increase, demonstrating the efficiency of the improved model design, its strong practicality, and its robustness.

[0122] Table 7 shows the detection results for the small object category "safety helmets," demonstrating a significant improvement in the model's ability to handle small objects. Introducing DPC-Block on Model A resulted in mAP@0.5 and mAP@0.5-0.95 results of 0.928 and 0.512, respectively, representing improvements of 4.4% and 2.2% compared to Model A. The recall result was 0.865, a 1.9% improvement over Model A, and the Prams score was 3.9M, a mere 0.9M increase compared to Model A. This demonstrates that DPC-Block effectively enhances small object detection capabilities and reduces missed detections while minimizing the need for additional parameters. Shape-IoU is further introduced on DPC-Block, and its mAP@0.5 and mAP@0.5-0.95 results are 0.933 and 0.537, respectively, which are improved by 0.5% and 2.5%, respectively, indicating the effectiveness of Shape-IoU for small target detection in complex situations. ModelC introduces the MLCA attention mechanism based on ModelB to enhance the model's understanding of details and global information. Its mAP@0.5 and mAP@0.5-0.95 results are 0.942 and 0.556, respectively, which are improved by 0.9% and 1.9%, respectively, indicating the effectiveness of the MLCA attention mechanism for small target detection in complex situations.

[0123] Compared to the overall category detection results, the performance improvement for small object detection is even more significant, demonstrating the importance of the introduced modules (particularly the DPC-Block) in capturing small object features and context. Furthermore, the final model design (Model D) approaches ideal values for Recal and Map@0.5, validating the adaptability and advantages of YOLOv8-DPMS for small object detection tasks in complex underground environments. Figure 17 The map@0.5 result graph of the ablation experiment for each category

[0124] This paper proposes a small target detection algorithm YOLOv8-DPMS, which is based on conventional convolution, C2f module and spatial pyramid pooling module, and integrates the MLCA attention mechanism module that focuses on channel, space and position information to construct a small target feature extraction network based on the MLCA attention mechanism. A small target feature fusion and enhancement module DPC-Block based on deformable convolution and residual structure is proposed, and the MLCA attention mechanism is integrated to fuse and enhance the shallow and deep features in the feature extraction network, thus constructing a feature fusion and enhancement network. Based on the above two networks, small target detection is achieved by combining the target classification and bounding box prediction modules. The experimental results on the target detection dataset show that:

[0125] The YOLOv8-DPMS algorithm proposed in this paper achieves mAP@0.5 and mAP@0.5-0.95 of 98.1% and 78.7% on the Coal-H dataset overall, representing improvements of 3.4% and 3.6% respectively compared to the YOLO series models. In the small object category of hardhat detection, mAP@0.5 and mAP@0.5-0.95 reach 94.2% and 55.6%, respectively, representing improvements of 10.8% and 12.4% respectively compared to the YOLO series models.

[0126] In summary, the YOLOv8-DPMS algorithm proposed in this paper can effectively improve the accuracy of small target detection.

[0127] Those skilled in the art will understand that the above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will still be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the scope of protection of the invention. All technical features in this embodiment may be freely combined according to actual needs.

[0128] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A mine small target detection method based on deformable convolution and residual structure, characterized by: Specifically include: (1) Image preprocessing module: preprocess the collected original target image; (2) Small target feature extraction network module based on MLCA attention mechanism: This module focuses on channel, spatial, and position information during the small target feature extraction process, suppresses background noise interference, and enhances the expressiveness of the extracted features. The preprocessed image is input into this module to suppress the interference of background noise in the small target image and extract features with high representational capabilities. The features at different layers are then input into the feature enhancement fusion module. (3) Small target feature fusion and reinforcement network module based on deformable convolution and residual structure: The different scale features output by the C2f module and the spatial pyramid pooling module in the feature extraction network are input into the feature fusion and reinforcement network, so that the shallow and deep features in the feature extraction network are fused and reinforced, while learning high-order semantic features, the rich detail information represented by the low-order features is retained, thereby effectively improving the accuracy of small target detection; (4) Target classification and bounding box prediction network: Based on the output of the four MLCA modules of the feature fusion enhancement network module, the prediction network can focus on important features, forming four target prediction branches, and using different anchor boxes to predict targets of different sizes. Each anchor box corresponds to a target of a specific size. During the training process, in order to speed up the convergence of the target detection model and improve the regression accuracy of the target prediction box, the ShapeIOU loss function is used to optimize the original loss function of the network to improve the detection accuracy.

2. The method for detecting small targets in mines based on deformable convolution and residual structure according to claim 1, characterized in that: The principle of the image preprocessing module specifically includes the following steps: using the Mosaic data augmentation method to process the original image. The Mosaic method uses four pictures and splices them in a random scaling, cropping and arrangement manner to combine multiple images. It is used to expand the data set while improving the model training speed and reducing memory requirements.

3. The method for detecting small targets in mines based on deformable convolution and residual structure according to claim 1, characterized in that: The small target feature extraction network based on the MLCA attention mechanism consists of 5 Conv, 4 C2f modules, 4 MLCA modules, and 1 spatial pyramid pooling module SPPF; Among them, the Conv module is a commonly used basic module, which consists of two-dimensional convolution Conv2d, batch normalization BN and activation function SiLU. It is used to enhance the nonlinear ability of convolution operation and ensures that the model can capture low-level features as a basic unit. During the forward propagation process, C2f splits the input feature map into two parts, performs convolution and Bottleneck processing on these two parts respectively. The processed feature maps are recombined and then pass through the final convolution layer to generate the output, which enhances the expressiveness of the feature map and improves the model's ability to recognize complex targets. The MLCA module adopts a multi-scale channel perception strategy, taking into account both channel information and spatial information. It combines local and global information to improve the network's feature extraction capability and further enhance the network's recognition accuracy for small targets. The SPPF module receives the feature map of the previous SPPF layer as input and uses multiple small pooling layers to replace the pooling operation of a single large core. Through pooling operations at different scales, SPPF can extract richer multi-scale features and generate a fixed-length feature vector by splicing the outputs of each pooling layer. This ultimately enhances the network's adaptability to targets of different sizes, thereby further improving the effectiveness of small target detection. Based on the Conv module, C2f module, MLCA module and SPPF module, a small target feature extraction network based on the MLCA attention mechanism is constructed. Through this structure, feature extraction of the image is performed, which can integrate attention to channel, space and position information into the process of small target feature extraction, better suppress the interference of background noise and enhance the expressiveness of the extracted features, and provide features with strong expressiveness for subsequent feature enhancement and fusion tasks.

4. The method for detecting small targets in mines based on deformable convolution and residual structure according to claim 1, characterized in that: The small target feature fusion enhancement network module based on deformable convolution and residual structure consists of 6 CDPC modules, 3 upsampling modules, 3 Conv modules, and 6 Concat modules; through dynamic feature fusion and cross-scale interaction, it effectively enhances the target detection capability in complex scenes.

5. The method for detecting small targets in mines based on deformable convolution and residual structure according to claim 1, characterized in that: The detection layer and loss function in the target classification and bounding box prediction network are improved as follows: the detection head structure and loss function are improved and designed, and in order to make the detection head pay more attention to important features and thus perform small target detection, the MLCA attention mechanism is added before detect; wherein, the improved detection head structure includes multiple convolutional layers Conv and a two-dimensional convolutional layer Conv2d, which is divided into a bounding box regression branch and a classification branch. The regression branch adopts a combined loss based on ShapeIOU and Distribution Focal Loss to optimize the bounding box positioning accuracy; the classification branch introduces binary cross entropy loss BCE to improve classification performance; ShapeIOU effectively solves the problem of gradient vanishing of traditional IoU in small overlapping areas, while Distribution Focal Loss further refines the predicted distribution of the bounding box, effectively improving the convergence time of the model.

Images