Coal mine underground intelligent equipment target detection method based on SPPF-Vannet and FPN-BGA

Abstract

The invention discloses a coal mine underground intelligent equipment target detection method based on an SPPF-Vannet and an FPN-BGA, belongs to the field of coal mine underground intelligent equipment target detection, and provides a small target detection model SPPF-Vannet-BGA based on the SPPF-Vannet and the FPN-BGA. IEM and Mosaic data enhancement methods are fused to pre-process images, so that the sample space is expanded, the training speed is improved, and the memory requirement is reduced; according to the SPPF-Vannet-based feature extraction network module, a VanilaNet module and an improved SPPF module are combined, so that interference of background noise in a small target image can be suppressed, and features with high characterization capability can be extracted; a small target feature fusion enhancement method FPN-BGA based on CSPGS and BIFPN is provided, the complex non-uniform feature extraction and global feature capture capability is improved, and network parameters and floating point calculation amount can be effectively reduced; based on the loss function of the SIOU, the convergence speed of the target detection model can be increased, the target detection precision can be improved, and the target detection effect in the underground complex environment can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology for intelligent equipment in underground coal mines, and particularly relates to a target detection method for intelligent equipment in underground coal mines based on SPPF-Vannet and FPN-BGA. Background Technology

[0002] Currently, coal continues to play a vital role in my country's energy security, accounting for over 50% of primary energy consumption. As mining activities advance deeper, the underground space exhibits complex characteristics of high stress, strong disturbance, and multi-field coupling, making traditional manual assistance methods insufficient for the development requirements of intelligent mining. Under the national strategy of promoting intelligent mining, mining robot technology is becoming a key breakthrough in reducing manpower, increasing efficiency, and improving inherent safety. Underground auxiliary robots and intelligent equipment undertake tasks such as material handling, equipment assembly, roadway support, dust control, and equipment maintenance. These equipment systems, relying on technologies such as visual perception, autonomous navigation, and intelligent decision-making, are gradually replacing manual labor. However, the underground working environment in coal mines suffers from multiple adverse factors, including uneven lighting, pervasive dust, and limited field of vision, leading to potential misjudgments in existing underground visual target detection technologies.

[0003] Currently, algorithms for underground target detection are mainly divided into two categories: single-stage and two-stage detection algorithms. Single-stage detection algorithms are known for their advantages such as high speed, simple structure, ease of deployment, and strong real-time performance. Classic examples include the YOLO (You Only Look Once) series and the SSD (Single Shot Detector) series. Two-stage detection algorithms, through a dual process of candidate box generation and fine-tuning of classification, offer greater advantages in target localization and classification accuracy, especially when handling multi-scale targets. Typical two-stage algorithms include R-CNN (Region with CNN feature) and Faster R-CNN, among others. However, two-stage algorithms are more complex and difficult to deploy, making them unsuitable for applications with high real-time requirements. Given the actual needs of current underground coal mine operations, real-time performance and rapid deployment are crucial indicators for target detection; therefore, single-stage target detection algorithms are more advantageous than two-stage algorithms. In recent years, with the continuous development of the YOLO network, many researchers have proposed target detection models based on YOLO suitable for different application scenarios. However, for small target detection scenarios, the algorithm still faces problems such as large number of model parameters, long inference time, and high false detection and false negative rates. Small target detection still faces severe challenges in achieving a balance between lightweight, real-time and high accuracy.

[0004] In summary, although significant progress has been made in the current single-stage small target detection model, further research is still needed on the following two aspects for its application in coal mine target detection: (1) How to effectively perform preprocessing operations such as enhancement and denoising on the acquired original images under the influence of complex environments to improve image recognizability; (2) 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; (3) Research on a single-stage target detection network that balances lightweight, real-time performance, and high detection accuracy. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the problems of large number of model parameters, long inference interaction time, and high false detection and false negative rates in current small target detection algorithms. It proposes a resource-programmable elastic acceleration method based on YARN, which effectively improves the ability to extract complex non-uniform features and capture global features, and can effectively reduce network parameters and floating-point computation, accelerate the convergence speed of the target detection model and improve the target detection accuracy.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA includes four modules: an image preprocessing module, a feature extraction network module based on SPPF-Vannet, a small target feature fusion enhancement network module based on CSPGS, ContMix, BIFPN and CA attention mechanism, and a target classification and bounding box prediction module.

[0008] The image preprocessing module is used to preprocess the acquired raw images containing the target. It combines the illumination estimation module (IEM) and the Mosaic data augmentation method to process the raw images. It stitches together a set of images by random scaling, cropping and arranging, thereby combining multiple images. This expands the dataset while improving the training speed of the model and reducing memory requirements.

[0009] The SPPF-Vannet-based feature extraction network module is constructed by integrating the fast spatial pyramid pooling module of the SEAM module on the basis of the VanillaNet module. The preprocessed image is input into this module to suppress the interference of background noise in small target images and extract features with high representation ability. Then, the features of different layers are input into the feature enhancement and fusion module.

[0010] The small target feature fusion enhancement network module based on CSPGS, ContMix, BIFPN and CA attention mechanism: It includes 4 CSPGS modules, 4 BiFPN convolutional modules, 2 ContMix modules, 1 CA attention mechanism module and 2 Upsample modules; it is used to improve the ability to extract complex non-uniform features and capture global features, and can effectively reduce network parameters and floating-point computation.

[0011] The target classification and bounding box prediction module consists of three CSPGS modules whose outputs form three target prediction branches based on the feature fusion enhancement network module. Different anchor boxes are used to predict targets of different sizes, with each anchor box corresponding to a target of a specific size. During training, in order to accelerate the convergence speed of the target detection model and improve the regression accuracy of the target prediction anchor boxes, the SIOU loss function is minimized to optimize and improve the detection accuracy.

[0012] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the image preprocessing module specifically includes:

[0013] By adding an illumination estimation module (IEM), the input image is enhanced in low light conditions. The enhanced image is then augmented using the Mosaic data augmentation method, which stitches together a set of images by randomly scaling, cropping, and arranging them to combine multiple images.

[0014] In computer vision, image generation can be simplified to a formula similar to Retinex theory: S = R⊙L, where S is the input image, R is the intrinsic reflectivity of the object, and L is the illumination component. The goal of Image Image Prediction (IEM) is to estimate the illumination L from a noisy S to achieve image brightness restoration and detail enhancement. IEM employs an iterative optimization approach, with input u... k via hot start function Get initial illumination in, Input CNN module C IEM Calculate the regularized gradient pass Optimized lighting t k+1 ; use t k+1 Perform pixel-by-pixel division with the original input y to obtain the intermediate image u for the next round. k+1 Repeat the above process until, after K iterations, the final illumination t is output. K and intermediate image u K .

[0015] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the feature extraction network module based on SPPF-Vannet includes 5 VanillaNet modules and 1 SPPF-SE module.

[0016] The network structure was enhanced by introducing a dedicated feature optimization module; the receptive field enhancement module SPPF-SE module was constructed by concatenating the SPPF module with the VanilaNet module and integrating the SEAM attention module into the SPPF module.

[0017] The SEAM attention module employs depthwise separable convolutions with residual connections, reducing the number of parameters and improving feature extraction capabilities through channel-wise computation. It fuses the outputs of convolutions at different depths using pointwise convolutions, and then integrates channel features through two fully connected layers to achieve information exchange between channels. An exponential function is used to process the output of the fully connected layers, expanding the output range from [0, 1] to [1, e], thus correlating the effective features of occluded and unoccluded targets. The attention weights generated by the SEAM attention module are multiplied by the original features.

[0018] Based on the Vanillanet and SPPF modules, SPPF-Vannet is constructed. This structure is used to extract features from images, enabling the integration of channel, spatial, and location information into the extraction of small target features. This better suppresses background noise interference and enhances the expressive power of the extracted features, providing highly expressive features for subsequent feature enhancement and fusion tasks.

[0019] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the feature fusion network module FPN-BGA based on CSPGS module, ContMix module, BiFPN module and CA attention mechanism is used to achieve more ideal feature fusion and model lightweighting effect at multiple scales.

[0020] It is built upon CSPGS, upsampling, Concat, CA attention mechanisms, and ContMix; its structure includes two branches:

[0021] The first branch contains two CSPGS modules, two upsampling modules, and two Concat modules. The Concat module is derived from BiFPN and uses learnable normalized weights to perform weighted summation on the input feature maps, thereby achieving the concatenation of feature maps FMs at different scales.

[0022] The second branch contains two CSPGS modules, two ContMix modules, two Concat modules, and one CA module. By introducing the CA attention mechanism in the first layer of branch 2, the channel weights are dynamically adjusted to highlight key features, thereby enhancing the model's feature discrimination ability and improving the target localization accuracy. At the same time, the ContMix module introduced in branch 2 fully utilizes the contextual priors provided by the overview network, thereby enhancing the ability to comprehensively capture local contextual information.

[0023] Based on these two branches, the outputs of the three CSPGS modules, namely CSPGS1, CSPGS2 and CSPGS3, will be input into SHead for defect localization and classification. By combining GSConv and ContMix in the multi-scale feature fusion network, features from different levels are effectively aggregated to improve detection performance, significantly improve the detection sensitivity of intelligent equipment targets, and reduce computational costs.

[0024] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the CSPGS module designs a cross-level partial network through a one-time aggregation method, which reduces the complexity of computation and network structure while maintaining sufficient accuracy, and includes the Conv module and the GSConv module.

[0025] Depthwise separable convolution (DSC) is based on conventional convolutional neural networks. It divides the computation into two parts: one part is a depthwise convolution applied to a single input channel, and the other part is a regular pointwise convolution with a kernel size of 1x1. The depthwise convolution is performed in a two-dimensional plane, and the number of convolution kernels is the same as the number of channels in the previous layer.

[0026] The number of feature maps after Depthwise channel-wise convolution is the same as the number of channels in the input layer, which cannot expand the feature maps; Convolution is performed independently on each channel of the input layer, which does not effectively utilize the feature information of different channels at the same spatial location; Pointwise convolution is needed to combine these feature maps to generate new feature maps.

[0027] Pointwise convolution is very similar to regular convolution. Its kernel size is 1×1×M, where M is the number of channels in the previous layer. Therefore, the convolution operation here will weight and combine the feature maps from the previous step in the depth direction to generate new feature maps. The number of output feature maps is equal to the number of convolution kernels.

[0028] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, in order to enable convolution to simultaneously adapt to different input resolutions and retain strong local inductive bias, and to make full use of the contextual priors provided by the overview network, a context-mixed dynamic convolution ContMix is ​​introduced. By calculating the affinity map between each token in the feature map and the center tokens of multiple regions, the connection between the token and the global context is represented. Then, the affinity map is converted into a dynamic convolution kernel in a learnable manner, and global contextual information is injected into each weight inside the convolution kernel. When the dynamic convolution kernel acts on the feature map through a sliding window, each token will be modulated with global information. When such a dynamic kernel acts on the feature map through a sliding window, each element in the feature map will be modulated by the approximate global information collected through the region center, thereby effectively modeling long-range dependencies.

[0029] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the Bidirectional Feature Pyramid Network (BiFPN) is used. BiFPN removes nodes with only one in-degree, eliminating some redundant computations; it adds several residual connections, where residual connections connect original feature nodes in the same layer to the output node to participate in bottom-up feature fusion, and constructs top-down and bottom-up fusion into a single module, repeatedly stacking them to enhance information fusion; deep features are continuously upsampled and fused with shallow features, and after upsampling, shallow features are continuously downsampled and fused with deep features; furthermore, the network adds two lateral connection paths to the original FPN+PAN structure to integrate the features extracted by the backbone network into the feature map to be detected.

[0030] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA in this invention, the Coordinate Attention (CA) mechanism decomposes channel attention into a feature encoding process that runs in parallel along two directions, horizontal and vertical. This enables the establishment of long-range dependencies in one direction while preserving positional information in the other. The two are then concatenated to obtain an attention map that has both direction awareness and positional information. Finally, after being processed by BN and Sigmoid functions, the attention map is applied to the input feature map to enhance the expressive power of the features.

[0031] Given an input X, each channel is encoded along the horizontal and vertical directions using pooling kernels of size (H,1) and (1,W), respectively. The output of the c-th channel at width w and height h can be expressed by the following formula:

[0032]

[0033] The two encoded feature maps are concatenated and then transformed using a 1×1 convolution function F1, as shown in the following formula:

[0034] f=δ(F1([z h ,z w ])) (4)

[0035] In the formula, f represents the feature mapping result of spatial information in the horizontal and vertical directions, δ refers to the nonlinear activation function, and [·,·] is the feature aggregation in space; f decomposition produces f h ∈R C / r×H and f w ∈R C / r×W Two independent features; using two 1×1 convolution functions F w and F h f respectively w and f h The number of channels is transformed to match the number of channels in the input X, as shown in formula (3):

[0036]

[0037] In the formula, σ is the Sigmoid function; g w and g h The expanded output formula of the CA module is as follows:

[0038]

[0039] As a further preferred embodiment of the target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA of the present invention, the expression for minimizing the SIOU loss function is as follows:

[0040]

[0041] Where IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box, Δ is the distance cost, and Ω is the shape cost.

[0042] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0043] 1. This invention proposes a feature extraction network SPPF-Vannet based on SPPF and VanillaNet. In order to suppress the interference of background noise in small target images and extract features with high representation ability, a fast spatial pyramid pooling module integrating the SEAM module is fused into the VanillaNet module to construct a feature extraction network.

[0044] 2. This invention proposes a small target feature fusion enhancement network module FPN-BGA based on CSPGS, ContMix, BIFPN, and CA attention mechanism. It can improve the extraction of complex non-uniform features and global feature capture capabilities, and effectively reduce network parameters and floating-point computation. To effectively extract and enhance the detailed features of small targets in images, it combines the advantages of shallow and deep features, retaining the rich detailed information represented by low-order features while learning high-order semantic features, thereby effectively improving the accuracy of small target detection. To maintain ideal detection performance while reducing network parameters and floating-point computation, a lightweight small target feature fusion enhancement network is constructed by introducing BiFPN, ContMix, CA attention mechanism modules, and Upsample. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the SPPF-Vannet-BGA algorithm structure of the present invention;

[0046] Figure 2 This is a schematic diagram of the IEM structure of the present invention;

[0047] Figure 3 This is a schematic diagram of the internal structure of the SPPF-SE module of this invention;

[0048] Figure 4 This is a structural diagram of the SPPF-Vannet of this invention;

[0049] Figure 5 This is a schematic diagram of the feature fusion network structure of the present invention;

[0050] Figure 6 This is a diagram of the internal structure of the CSPGS module of this invention;

[0051] Figure 7 This is a flowchart of the Depthwise convolution workflow of the present invention;

[0052] Figure 8 This is a flowchart of the Pointwise convolution workflow of the present invention;

[0053] Figure 9 This is a structural diagram of the GSConv of the present invention;

[0054] Figure 10 This is a schematic diagram of the ContMix structure of the present invention;

[0055] Figure 11 This is the FPN feature pyramid of the present invention;

[0056] Figure 12 This is the PAN feature pyramid of the present invention;

[0057] Figure 13This is a diagram of the BiFPN structure of the present invention;

[0058] Figure 14 This is a diagram of the BiFPN Layer network structure of the present invention;

[0059] Figure 15 This is a schematic diagram of the feature pyramids of FPN (left), PAN (middle), and BiFPN (right) of the present invention;

[0060] Figure 16 This is a structural diagram of the CA attention module of the present invention;

[0061] Figure 17 This is a schematic diagram of the angle loss of the present invention;

[0062] Figure 18 This is a schematic diagram of the relevant parameters of the SIoU loss function of this invention;

[0063] Figure 19 This is a schematic diagram of the intelligent analysis system architecture for the equipment operation status of the present invention;

[0064] Figure 20 This is a schematic diagram illustrating an example of a self-built dataset from this invention;

[0065] Figure 21 This is a schematic diagram showing the detection results of the present invention;

[0066] Figure 22 This is a schematic diagram showing the detection results of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0069] This invention proposes a small target detection model, SPPF-Vannet-BGA, based on SPPF-Vannet and FPN-BGA. It integrates illumination estimation (IEM) and Mosaic data augmentation methods for image preprocessing, expanding the sample space, improving training speed, and reducing memory requirements. A feature extraction network module based on SPPF-Vannet is designed, combining the VanillaNet module with an improved SPPF module, which can suppress background noise interference in small target images and extract highly representative features. A small target feature fusion enhancement method, FPN-BGA, based on CSPGS and BIFPN, is proposed, improving the extraction of complex non-uniform features and global feature capture capabilities, while effectively reducing network parameters and floating-point computation. Finally, a loss function based on SIOU is designed to accelerate the convergence speed of the target detection model and improve target detection accuracy. Experimental results show that the SPPF-Vannet-BGA model can effectively improve target detection performance in complex downhole environments.

[0070] This invention focuses on feature extraction based on convolutional neural networks, attention mechanism embedding, multi-scale feature fusion, and model lightweighting. It proposes a target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA, denoted as SPPF-Vannet-BGA. The main contributions are as follows:

[0071] (1) A feature extraction network SPPF-Vannet based on SPPF and VanillaNet is proposed. In order to suppress the interference of background noise in small target images and extract features with high representation ability, a fast spatial pyramid pooling module integrating the SEAM module is fused into the VanillaNet module to construct the feature extraction network.

[0072] (2) A small target feature fusion enhancement network module FPN-BGA based on CSPGS, ContMix, BIFPN, and CA attention mechanism is proposed. This module can improve the extraction of complex non-uniform features and global feature capture capabilities, and effectively reduce network parameters and floating-point computation. To effectively extract and enhance the detailed features of small targets in images, the advantages of shallow and deep features are combined. While learning high-order semantic features, the rich detailed information represented by low-order features is preserved, thereby effectively improving the accuracy of small target detection. To maintain ideal detection performance while reducing network parameters and floating-point computation, a lightweight small target feature fusion enhancement network is constructed by introducing BiFPN, ContMix, CA attention mechanism modules, and Upsample.

[0073] To address the issues of large model parameters, long inference interaction time, and high false positive and false negative rates in current small object detection algorithms, this paper investigates feature extraction based on VanillaNet, attention mechanism embedding, bidirectional feature pyramid fusion, and model lightweighting. A fast small object detection algorithm based on an improved VanillaNet and the feature fusion network module FPN-BGA is proposed. Figure 1 As shown, the algorithm consists of four modules:

[0074] (1) Image preprocessing module. This module preprocesses the acquired raw images containing the target, combines the illumination estimation module IEM and the Mosaic data augmentation method to process the raw images, and stitches together a set of images in a random scaling, cropping and arrangement manner to combine multiple images. This can improve the training speed of the model while expanding the dataset and reducing memory requirements.

[0075] (2) Feature extraction network module based on SPPF-Vannet. This module consists of 5 VanillaNet modules and 1 SPPF-SE module. The preprocessed image is input into this module to suppress the interference of background noise in small target images and extract features with high representation ability. Then, the features from different layers are input into the feature enhancement and fusion module.

[0076] (3) A small target feature fusion enhancement network module based on CSPGS, ContMix, BIFPN and CA attention mechanism is used to improve the ability to extract complex non-uniform features and capture global features, and can effectively reduce network parameters and floating-point computation. This module consists of 4 CSPGS modules, 4 BiFPN convolutional modules, 2 ContMix modules, 1 CA attention mechanism module and 2 Upsample modules.

[0077] To address the complex and non-uniform feature information often found in downhole image data, this invention replaces the C2f module in the original network with the CSPGS module and introduces GSConv to fuse group convolution and depthwise separable convolution (DWConv) mechanisms. Through collaborative optimization, computational efficiency is improved and computational costs are reduced. To address the problem of limited receptive field leading to weak comprehensive capture of local contextual information, this invention also introduces a dynamic convolution module, ContMix, to replace the Conv module in the original network.

[0078] (4) Target classification and bounding box prediction module. The outputs of the three CSPGS modules based on the feature fusion enhancement network module constitute three target prediction branches, which use different anchor boxes to predict targets of different sizes. Each anchor box corresponds to a target of a specific size. During training, in order to accelerate the convergence speed of the target detection model and improve the regression accuracy of the target prediction anchor boxes, the SIOU loss function is minimized to optimize and improve the detection accuracy.

[0079] Image Preprocessing Module: In coal production, the prevention and control of safety accidents are crucial. However, the underground environment of coal mines is complex and variable, often accompanied by poor lighting conditions, high dust concentration, and high humidity. These factors affect the stability of visual monitoring and place higher demands on the generalization ability of target detection algorithms. To address these issues, this study designed an image preprocessing module. By incorporating an Illumination Estimation (IEM) module, the input images are enhanced under low-light conditions. The enhanced images are then augmented using the Mosaic data augmentation method, which stitches together a set of images by random scaling, cropping, and arrangement. This combination of multiple images expands the dataset while improving the model's training speed and reducing memory requirements.

[0080] In computer vision, image generation can be simplified to a formula similar to Retinex theory: S = R⊙L, where S is the input image, R is the intrinsic reflectivity of the object, and L is the illumination component. The goal of Image Image Prediction (IEM) is to estimate the illumination L from the noisy S to achieve image brightness restoration and detail enhancement.

[0081] IEM employs an iterative optimization approach, with input u k via "hot start function" "Obtain initial illumination" in Input CNN module C IEM Calculate the regularized gradient ";pass Optimized lighting t k+1 ; use t k+1 Perform pixel-by-pixel division with the original input y to obtain the intermediate image u for the next round. k+1 Repeat the above process until, after K iterations, the final illumination t is output. K and intermediate image u K .

[0082] The Illumination Estimation (IEM) module can avoid interference from "overexposed pixels" while enhancing low-light images. By replacing manually designed regularization terms with neural networks, it can more flexibly handle complex lighting scenarios such as the dim light in coal mines and the strong reflections in industrial scenes.

[0083] The SPPF-Vannet feature extraction network is based on SPPF and VanillaNet. SPPF-Vannet consists of five VanillaNet modules and one SPPF-SE module. VanillaNet contains only basic convolutional and pooling layers, without complex or skip connections, thus reducing the number of model parameters. Its concise structure and deep training strategy achieve a balance between parameter control and basic performance. However, when detection tasks involve diverse targets and cluttered backgrounds, especially when targets are occluded to varying degrees, this network, lacking a targeted feature enhancement mechanism, struggles to effectively capture and focus on key target information, leading to decreased feature recognition and a high false negative rate. To improve adaptability to complex occlusion scenarios while maintaining the model's lightweight advantage, this invention enhances the network structure by introducing a dedicated feature optimization module. Here, the SPPF module is concatenated with VanilaNet, and the SEAM (spatially enhanced attention module) module is integrated into the SPPF module to construct the receptive field enhancement module SPPF-SE. This improves the model's adaptability to complex occlusion scenes while maintaining its lightweight advantage, thereby improving the detection accuracy of occluded targets. Its structure is as follows: Figure 3 As shown.

[0084] SEAM attention module design such as Figure 3 As shown, firstly, a depthwise separable convolution with residual connections is employed to reduce the number of parameters and improve feature extraction capabilities through channel-wise computation. To enhance the feature correlation between channels, pointwise convolutions are used to fuse the outputs of convolutions at different depths. Subsequently, a two-layer fully connected network integrates channel features to achieve information interaction between channels. Here, an exponential function is used to process the output of the fully connected layer, expanding the output value range of the fully connected layer from [0,1] to [1,e], enabling the effective features of occluded and unoccluded targets to be correlated. Finally, the attention weights generated by the SEAM module are multiplied with the original features, effectively mitigating the feature loss problem caused by occlusion.

[0085] Based on the Vanillanet and SPPF modules, this paper constructs SPPF-Vannet, whose structure is as follows: Figure 4 As shown in the figure. By using this structure to extract features from images, it is possible to incorporate attention to channel, spatial, and positional information during the extraction of small target features, better suppress the interference of background noise, and enhance the expressive power of the extracted features, thus providing highly expressive features for subsequent feature enhancement and fusion tasks.

[0086] Feature fusion network module based on CSPGS, ContMix, BiFPN, and CA attention mechanism (FPN-BGA)

[0087] Conventional feature fusion networks are typically composed of Figure 5 (a) shows a network consisting of an upsampling module, Concat, C2f, and Conv. However, conventional feature fusion networks can only learn low-level features with rich details (such as edges and textures) through shallow neurons in convolutional neural networks. As the number of layers increases, the accumulated feature space becomes larger, and the receptive field of the corresponding feature map also becomes larger. Although it can learn higher-level features, it also causes the loss of a lot of detailed information, resulting in a decrease in the accuracy of small object detection.

[0088] To address the complex and non-uniform feature information often found in downhole image data, this invention replaces the C2f module in the original network with the CSPGS module and introduces GSConv to fuse group convolution and depthwise separable convolution (DWConv) mechanisms. Through collaborative optimization, computational efficiency is improved and computational costs are reduced. To address the problem of limited receptive field leading to weak comprehensive capture of local contextual information, this invention also introduces a dynamic convolution module, ContMix, to replace the Conv module in the original network.

[0089] To address the aforementioned issues, this study designs a novel multi-scale feature fusion network, FPN-BGA, constructed using CSPGS, ContMix, BiFPN, and CA attention mechanisms. This network achieves more ideal feature fusion and lightweight model implementation across multiple scales, effectively extracting and enhancing the detailed features of small targets in images, thereby improving the accuracy of small target detection. Figure 5As shown in (b), FPN-BGA is constructed based on CSPGS, upsampling, Concat, CA attention mechanism, and ContMix. Its structure comprises two branches: the first branch contains two CSPGS modules, two upsampling modules, and two Concat modules. The Concat module, derived from BiFPN, uses learnable normalized weights to perform weighted summation of the input feature maps, thereby achieving the concatenation of feature maps (FMs) at different scales. The second branch contains two CSPGS modules, two ContMix modules, two Concat modules, and one CA module. By introducing the CA attention mechanism in the first layer of branch 2, the channel weights are dynamically adjusted to highlight key features, thereby enhancing the model's feature discrimination ability and improving target localization accuracy. Simultaneously, the ContMix module introduced in branch 2 fully utilizes the contextual priors provided by the overview network, thereby enhancing the comprehensive capture of local contextual information. Based on these two branches, the outputs of the three CSPGS modules (CSPGS1, CSPGS2, and CSPGS3) are used to input the SHead for defect localization and classification. By combining GSConv and ContMix in a multi-scale feature fusion network, it can effectively aggregate features from different levels to improve detection performance, significantly improve the detection sensitivity of intelligent equipment targets, and reduce computational costs.

[0090] The CSPGS module: The CSPGS module designs cross-level partial networks using a one-time aggregation method, reducing computational and network structure complexity while maintaining sufficient accuracy. It mainly consists of the Conv module and the GSConv module, such as... Figure 6 As shown.

[0091] Depthwise Separable Convolution (DSC) is based on conventional convolutional neural networks, dividing the computation into two parts: a depthwise convolution applied to a single input channel, and a regular pointwise convolution with a kernel size of 1x1. The depthwise convolution is performed in a two-dimensional plane, and the number of kernels is the same as the number of channels in the previous layer.

[0092] The number of feature maps generated by depthwise channel-wise convolution is the same as the number of channels in the input layer, which cannot expand the feature maps. Moreover, this operation performs convolution operations independently on each channel of the input layer, failing to effectively utilize the feature information of different channels at the same spatial location. Therefore, pointwise convolution is needed to combine these feature maps to generate new feature maps.

[0093] Pointwise convolution is very similar to regular convolution. Its kernel size is 1×1×M, where M is the number of channels in the previous layer. Therefore, this convolution operation weights and combines the feature maps from the previous step along the depth direction to generate new feature maps. The number of output feature maps corresponds to the number of convolution kernels.

[0094] The feature map is extracted by combining depthwise (DW) and pointwise (PW) components. Compared to conventional convolution operations, depthwise separable convolution (DSC) has lower parameter count and computational cost. However, object detection is a challenging downstream task in computer vision, where speed and accuracy are equally important. Building a lightweight model with only a large number of depthwise separable convolutional layers cannot achieve sufficient accuracy. Because DSC separates the channel information of the input image during computation, its feature extraction and fusion capabilities are much lower than those of SC. GSConv uses shuffling to permeate the information generated by SC (channel-dense convolution operation) into all parts of the information generated by DSC (depthwise separable convolution). Shuffling is a uniform mixing strategy. This method allows the information from SC to be completely mixed into the output of DSC by uniformly exchanging local feature information on different channels, without cumbersome steps, making the output of DSC as close as possible to SC. Figure 9 This is a structural diagram of GSConv.

[0095] Dynamic Convolution with Context-Mixing (ContMix): To enable convolutions to adapt to different input resolutions and preserve strong local inductive biases, while fully utilizing the contextual priors provided by the overview network, dynamic convolution with context-mixing (ContMix) is introduced. Its core lies in representing the relationship between each token and the global context by calculating the affinity map between each token in the feature map and the center tokens of multiple regions. This affinity map is then transformed into a dynamic convolutional kernel in a learnable manner, and global contextual information is injected into each weight within the kernel. When the dynamic convolutional kernel operates on the feature map through a sliding window, each token is modulated with global information. When such a dynamic kernel operates on the feature map through a sliding window, each element in the feature map is modulated by approximate global information collected through the region centers, thus effectively modeling long-range dependencies.

[0096] like Figure 10 As shown, given the input feature map x∈ C ×H×W is first converted into two parts: W at the 1×1 convolutional layer.q Remodeling (Re) yields Q∈ C×HW =Re(W q X) and then aggregated into S×S region centers through adaptive average pooling, and then passed through a 1×1 convolutional layer W k Processed and reshaped to obtain Subsequently, the channels of Q and K were evenly divided into G groups, resulting in... and Because each pair of Q g and K g Having been flattened into a two-dimensional matrix, G affinity maps can be calculated using simple matrix multiplication:

[0097] To generate more robust feature representations, we define G spatially varying K×K dynamic kernels, as follows:

[0098] 1) Utilizing learnable linear layers Through A g With W d Matrix multiplication, aggregating each affinitymap A g The values ​​are stored in matrix row format (to save computation, all G affinity maps share the same W). d );

[0099] 2) Normalization is performed using the Softmax function;

[0100] 3) The above two steps can be expressed as: D g Each row is reshaped into the target kernel shape, thus generating a dynamic kernel that depends on the input at each token position;

[0101] 4) During convolution operations, the features are... Figure X The channels are also evenly divided into G groups, and channels in the same group share the same dynamic core.

[0102] 2.4.3 Bidirectional Feature Pyramid Network (BiFPN)

[0103] Feature Pyramid Network (FPN) is a multi-scale feature fusion method based on deep convolutional networks, used to improve the performance of object detection and image segmentation. FPN enhances the model's ability to detect objects at different scales by fusing features from different scales. Its core idea is to achieve more accurate object recognition by fusing multiple layers of feature maps, utilizing the semantic information of deep features and the detailed information of shallow features.

[0104] Because FPN integrates high-level features from deep convolutional layers and low-level features from shallow convolutional layers, it has excellent detection performance for small targets. The FPN structure is as follows: Figure 11 As shown.

[0105] PAN (Path Aggregation Network) is a deep neural network architecture for image semantic segmentation. PAN aggregates feature maps from different levels, ensuring that information from each feature map is fully utilized, thereby improving detection accuracy. It builds upon FPN by adding a bottom-up path. Since higher-level feature maps have stronger semantic information, which is beneficial for object classification, and lower-level feature maps have stronger positional information, which is beneficial for object localization, the predicted feature maps possess both high semantic and positional information, significantly improving object detection accuracy. PAN excels at using attention mechanisms to capture global information and model long-range dependencies, making it suitable for visual tasks requiring high global information. However, its performance is less ideal for multi-scale feature fusion tasks. The PAN structure is as follows: Figure 12 As shown.

[0106] BiFPN improves upon PANet's shortcomings, making it more suitable for multi-scale feature fusion tasks. BiFPN removes nodes with only one in-degree, eliminating redundant computations. It adds residual connections, connecting original feature nodes in the same layer to the output node for bottom-up feature fusion. Top-down and bottom-up fusion are constructed as a single module, allowing for repeated stacking and enhanced information fusion. Deep features are continuously upsampled and fused with shallow features; after upsampling, shallow features are continuously downsampled and fused with deep features. Furthermore, the network adds two lateral connection paths to the original FPN+PAN structure to integrate features extracted from the backbone network into the target feature map. This improvement allows for the fusion of more feature information in the target feature map without significantly increasing computation, enhancing the expressive power of the feature map and improving detector performance. Figure 13 The network structure diagram for adding BiFPN.

[0107] The calculation expression is:

[0108]

[0109]

[0110] The Resize operation is typically a downsampling or upsampling operation; w is a learned parameter used to distinguish the importance of different features during feature fusion. The BiFPN Layer network structure is as follows: Figure 14 As shown.

[0111] BiFPN, FPN, and PAN structures are compared to... Figure 15 As shown in the figure. It has three main improvements over FPN and PAN: (1) Removing single-in-degree nodes. Because these nodes do not perform feature fusion and contain relatively little information, they do not play a role in subsequent fusion. Removing them can result in lower computational cost, making the structure more lightweight and efficient. (2) Adding residual connections. For nodes at the same level, a connection is added between the original node that did not participate in feature fusion and the output node that participated in bottom-up fusion. This increases computational cost in exchange for richer fusion feature information. (3) Constructing a bidirectional fusion structure. The bidirectional path is built as a module, which can be stacked multiple times to achieve higher-level feature fusion.

[0112] 2.4.4 Coordinate Attention Mechanism

[0113] The CA (Carry-Ahead) mechanism innovatively embeds location information into the channel attention module, enhancing the fusion of spatial and channel-dimensional information in a novel way. By introducing coordinate information, the CA attention mechanism embeds location information into channel relationships, accurately encoding long-range contextual relationships. When processing images with complex backgrounds or small targets, it not only effectively captures horizontal and vertical spatial information but also dynamically allocates weights to each channel, improving information transmission efficiency and reducing interference from redundant information. Specifically, the CA module decomposes channel attention into parallel feature encoding processes along two directions, horizontal and vertical, establishing long-range dependencies in one direction while preserving location information in the other. These two processes are then concatenated to obtain an attention map that is both direction-aware and location-information-rich. Finally, after operations using Batch Normalization (BN) and the Sigmoid function, this map is applied to the input feature map to enhance feature expressiveness. Its structure is as follows: Figure 16 As shown.

[0114] Given an input X, each channel is encoded along the horizontal and vertical directions using pooling kernels of size (H,1) and (1,W), respectively. The output of the c-th channel at width w and height h can be expressed by the following formula:

[0115]

[0116] After the above transformation, not only is precise location information embedded, but a larger receptive field is also obtained. The two encoded feature maps are then concatenated and transformed using a 1×1 convolution function F1, as shown in the following formula:

[0117] f=δ(F1([z h ,z w ])) (4)

[0118] In the formula, f represents the feature mapping result of spatial information in the horizontal and vertical directions, δ refers to the nonlinear activation function, and [·,·] is the feature aggregation in space. Then, f decomposition produces fa. h ∈R C / r×H and f w ∈R C / r×W Two independent features. Finally, two 1×1 convolution functions F are used. w and F h f respectively w and f h The number of channels is transformed to be consistent with the number of channels in the input X, as shown in formula (3).

[0119]

[0120] In the formula, σ is the Sigmoid function. Let g... w and g h The expanded output formula of the CA module is as follows.

[0121]

[0122] In small object detection tasks, the CA attention mechanism can help the model better focus on the detailed features of small objects, thereby improving detection accuracy. In image classification tasks, the CA attention mechanism can enhance the model's attention to important features and improve classification accuracy.

[0123] 2.5 Improved LOSS (SIOU)

[0124] The effectiveness of object detection largely depends on the definition of the loss function. Traditional object detection loss functions rely on the aggregation of bounding box regression metrics, such as the distance between the predicted and ground truth boxes, the overlapping area, and the aspect ratio. This patent introduces the SIOU loss, which, after considering these regression metrics, also takes into account the vector angle between the required regressions. This allows the predicted box to quickly move to the nearest axis, effectively reducing the total degrees of freedom of the loss and significantly improving training speed and accuracy. The SIOU loss function consists of four cost functions: Angle cost, Distance cost, Shape cost, and IoU cost.

[0125] (1)Angle cost

[0126] Adding this angle-aware LF component minimizes the number of distance-dependent variables; the model attempts to first bring the predictions to the X or Y axis and then continue approaching along the relevant axis. Angle loss is as follows: Figure 17 As shown.

[0127] If α ≤ Π / 4, the convergence process will first minimize α; otherwise, it will minimize β.

[0128]

[0129] To achieve this first, use the following method to import and define LF components:

[0130]

[0131] in,

[0132]

[0133] (2) Distance cost

[0134] Distance cost was redefined based on Angle cost:

[0135]

[0136] in,

[0137]

[0138] As can be seen from the above formula, the contribution of distance cost decreases significantly when α approaches 0. Conversely, the closer α is to π / 4, the greater the contribution of distance cost. The problem becomes increasingly difficult as the angle increases. Therefore, γ is assigned a time-priority distance value as the angle increases.

[0139] (3) Shape cost

[0140] Shape cost is defined as:

[0141]

[0142] in,

[0143]

[0144] The value of θ defines the shape cost for each dataset, and its value is unique. The value of θ is a very important term in this equation, as it controls the degree of attention paid to the shape cost. If the value of θ is set to 1, it will immediately optimize a shape, thereby impairing the free movement of the shape.

[0145] (4) IoU Cost

[0146] The definition of IoU

[0147]

[0148] (5) SIOU Loss

[0149] Based on the above Angle cost, Distance cost, Shape cost, IoU cost, and SIOU loss function:

[0150]

[0151] The total loss function is:

[0152] L=W box L box +W cls L cls (18)

[0153] The relevant parameters of SIOU are as follows: Figure 18 As shown.

[0154] Experimental Verification: Example 1: To ensure the operational safety and continuous stable operation of the intelligent system for rapid conveyor belt recovery in underground coal mines, an intelligent monitoring and analysis system for equipment operation status based on machine vision and multi-dimensional sensor data was developed. Based on data access from multiple vehicle-mounted cameras, a target detection algorithm based on SPPF-Vannet-BGA and FPN-BGA was deployed to complete the development of an intelligent equipment status identification and early warning system. When abnormal operation status (conveyor failure, belt breakage, fatigue) or illegal intrusion is detected, the system automatically adjusts the equipment operation status according to set rules and operational logic to ensure the inherent safety of equipment operation and environmental safety. The system architecture diagram is as follows. Figure 19 As shown.

[0155] (1) Intelligent Detection of Belt Recycling Status: The intelligent system for rapid belt recycling in underground coal mines is equipped with a belt winding device. This device synchronously winds up the belt during the recycling process, completing the winding of the cut belt. If any abnormal conditions occur during the conveying, winding, or belt breakage processes, such as longitudinal or transverse bending, kinking, misalignment, tilting, or incomplete belt breakage, the intelligent control system needs to be activated to adjust the operating parameters and status of the belt recycling equipment to prevent accidents. To achieve intelligent detection and analysis of belt winding status, video cameras are installed on the conveying side, belt breakage side, and winding side of the automated belt recycling equipment. A video access system, the SPPF-Vannet-BGA target detection algorithm based on SPPF-Vannet and FPN-BGA, and an intelligent monitoring and control application system are deployed in the main control system. Combined with the belt winding operation specifications, abnormal detection is achieved during the belt conveying process and the winding process of the winding device. When abnormal conditions occur during the conveying, winding, or belt breakage processes, the status is automatically adjusted, and manual intervention is alerted.

[0156] (2) Intelligent detection and early warning of intrusion into the work area

[0157] The intelligent system for rapid recovery of conveyor belts in underground coal mines is designed to operate at the conveyor belt recovery points in various conveyor roadways. This presents challenges due to the complex environment, confined and variable spaces, and limited personnel movement in the underground roadways. Furthermore, the system's operation involves dangerous actions such as rotation, shaping, cutting, and conveying. Unauthorized entry of personnel or other moving targets into the operating area poses a safety hazard. To address this, an intelligent detection and early warning model for personnel intrusion within the equipment's operating area was developed. Cameras are installed at the front and rear ends of the intelligent system to acquire visual perception information of the roadway operating environment before and after the equipment. Visual feature analysis methods are used to obtain comprehensive visual feature information within the equipment's operating area. A detection algorithm based on SPPF-Vannet-BGA and FPN-BGA is deployed to improve the accuracy and generalization ability of the intelligent detection model, ensuring the safety of underground workers.

[0158] (3) Introduction to the dataset

[0159] The accuracy of the target recognition and detection model was not validated. Therefore, a coal mine underground video dataset was used to validate the model. This dataset is based on monitoring video data from various scenarios in real underground coal mine belt conveyor roadways. In these scenarios, equipment information is relatively fixed, the positions of facilities such as conveyor belts, tracks, and pipelines are fixed, and the camera installation points are also relatively stable. To construct the dataset, frames were first extracted from video clips, with one frame selected every 30 frames as a data sample. These extracted samples were labeled using the LabelImg tool, defining underground target categories such as pipelines, tracks, personnel, helmets, clothing, and belt areas. The annotation results were saved in XML and YOLO formats, forming files with annotation information. Figure 20 As shown, the self-built dataset contains 2000 experimental data points, covering a variety of typical downhole scenarios, ensuring sufficient data diversity and representativeness. The dataset is divided in an 8:2 ratio, with 1600 images used for training and 400 images used for testing.

[0160] (4) Comparative experiment: Figure 21This paper showcases the detection performance of the SPPF-Vannet-BGA network on this dataset. Through a visual comparison of the original samples, ground truth annotations, and prediction results, the detection performance of seven different targets is demonstrated. Experimental results show that the SPPF-Vannet-BGA network can still achieve accurate localization and classification when dealing with challenging scenarios involving complex morphological changes, low-contrast features, and significant irregularities in size and spatial distribution patterns, with good alignment of its predicted bounding boxes. These findings collectively validate the stability and effectiveness of this network in handling diverse target detection tasks for underground intelligent equipment. To rigorously verify the performance advantages of SPPF-Vannet-BGA in target detection of intelligent equipment in coal mines, this study conducts a comprehensive comparative analysis with cutting-edge detection algorithms. Table 1 presents the experimental results.

[0161] According to the data in Table 1, the proposed SPPF-Vannet-BGA detector demonstrates excellent detection accuracy across all seven target categories, with category-specific average accuracies of 99.6%, 99.20%, 97.80%, 94.30%, 97.90%, 99.60%, and 98.70%, respectively. Particularly noteworthy is its mAP@0.5 (average accuracy at IoU = 0.5), reaching 98.20%. While the two-stage algorithm Faster-RCNN performs best in detecting large targets (such as conveyor belts), achieving a detection accuracy of 1.00, demonstrating its strong performance in large target detection, its model has a large number of parameters, making it less ideal under limited computational resources. Similarly, although the SPPF-Vannet-BGA algorithm achieves an accuracy of 0.97 in pipe detection, its high computational complexity may affect the efficiency of real-time applications. In contrast, SSD and YOLOX perform poorly in detecting small targets (such as helmets and clothing) and exhibit false negatives, indicating that these algorithms struggle with multi-target detection tasks in complex scenes. Overall, the SPPF-Vannet-BGA algorithm performs exceptionally well in multi-target detection in underground coal mines, particularly in handling small targets and complex environments. Compared to other algorithms, this method not only boasts higher accuracy and greater adaptability but also demonstrates significant practical value in real-world applications.

[0162] Table 1

[0163]

[0164] (5) Ablation Experiment: To verify the effectiveness of SPPF-Vannet-BGA, further ablation experiments were conducted. For this purpose, SPPF-Vannet, FPN-BGA and SHead models were used for ablation verification. To present the ablation experiment results intuitively, the five ablation schemes were simplified to the following abbreviations: (1) Baseline is called BL, and its loss function is CIoU; (2) Baseline + SIoU loss function is called BLS; (3) Baseline + SIoU loss function + SPPF-Vannet is called BLS-SV; (4) SPPF-Vannet-BGA detection algorithm model proposed in this invention.

[0165] Analysis of the data in Table 2 shows that significant performance improvements were achieved through continuous architecture optimization. Preliminary comparisons indicate that the detection performance of the baseline BL model is significantly inferior to other variants. Specifically, after integrating the enhanced detection head (SHead) into BLS, mAP@0.5 increased from 96.3% (BL) to 97.5%, a 1.2% improvement, and mAP@0.5-0.95 increased from 76.2% (BL) to 77.1%, a 0.9% improvement, while recall increased by 0.2%, confirming the effectiveness of the optimized detection head design. Subsequently, introducing the SPPF-Vannet feature extraction network into BLS-SV further improved mAP@0.5 to 97.70%, mAP@0.5-0.95 by 0.9%, and recall by 1.2%, fully demonstrating the performance advantages of SPPF-Vannet-BGA in target detection for intelligent equipment in underground coal mines. Finally, the FPN-BGA feature fusion network was used to enhance the BLS-FV-FB model, resulting in a small but significant improvement in model performance.

[0166] Table 2

[0167]

[0168] Example 2: (1) Introduction to the dataset

[0169] To verify the applicability of the algorithm model proposed in this patent to other fields, a defect recognition verification experiment was conducted on the NEU-DET steel surface defect dataset using the proposed SPPF-Vannet-BGA model. The NEU-DET dataset is a publicly available image library of steel surface defects provided by Northeastern University, containing six different types of steel surface defects: inclusions (In), cracks (Cr), patches (Pa), rolled oxide scale (Rs), pitting (Ps), and scratches (Sc). This dataset consists of 1800 grayscale images (300 images per defect type), with a pixel size of 200×200. To support the training, validation, and testing phases of the detection model, 1778 images were selected from this dataset and divided into three parts: 356 images for testing, 178 images for validation, and 1244 images for training.

[0170] (2) Comparative Experiment

[0171] Figure 22 This paper showcases the detection performance of the SPPF-Vannet-BGA network on the NEU-DET dataset. Through a visual comparison of the original samples, ground truth annotations, and prediction results, the detection effectiveness for six different defect categories is demonstrated. Experimental results show that the SPPF-Vannet-BGA network can still achieve accurate defect localization and classification, with good bounding box alignment, even in challenging scenarios involving complex morphological changes, low-contrast features, and significant irregularities in defect size and spatial distribution patterns. These findings collectively validate the robustness and effectiveness of this network in handling diverse industrial defect detection tasks. To rigorously verify the performance advantages of SPPF-Vannet-BGA in steel surface defect detection, this study conducts a comprehensive comparative analysis with cutting-edge detection algorithms. Table 1 presents the experimental results.

[0172] Table 3

[0173]

[0174]

[0175]

[0176] According to the data in Table 3, the proposed SPPF-Vannet-BGA detector exhibits excellent detection accuracy across all six defect categories, with category-specific average accuracies of 97.60%, 92.20%, 99.40%, 96.30%, 96.70%, and 98.20%, respectively. Particularly noteworthy is its mAP@0.5 (average accuracy at IoU = 0.5) reaching 96.70%, a significant improvement of 16.00-26.20% compared to the baseline.

[0177] (3) Ablation test

[0178] To verify the effectiveness of SPPF-Vannet-BGA, this study conducted ablation experiments on the NEU-DET dataset. For this purpose, we used our self-developed SPPF-Vannet, FPN-BGA, and SHead models for ablation verification. To present the ablation experiment results intuitively, this study simplified the five ablation schemes into the following abbreviations: (1) Baseline is called BL with a loss function of CIoU; (2) Baseline + SIoU loss function is called BLS; (3) Baseline + SIoU loss function + SPPF-Vannet is called BLS-SV; (4) Baseline + SIoU loss function + SPPF-Vannet + FPN-BGA is called BLS-SV-FB, i.e., the SPPF-Vannet-BGA network.

[0179] The ablation experiments on the NEU-DET dataset are shown in Table 4. The results demonstrate that significant performance improvements were achieved through continuous architecture optimization. Preliminary comparisons show that the detection performance of the Baseline variant is significantly inferior to other variants. Specifically, integrating the enhanced detection head (SHead) into BLS improved mAP@0.5 from 74.00% (BL) to 76.40%, an absolute gain of 2.4%, confirming the effectiveness of the optimized detection head design. Subsequently, introducing the SPPF-Vannet feature extraction network into BLS-SV resulted in a qualitative leap in performance, with mAP@0.5 increasing to 96.30%. This 19.9% ​​absolute gain (from 76.40% to 96.30%) fully demonstrates the superior ability of SPPF-Vannet-BGA in enhancing the feature representation of steel defect detection. Finally, the FPN-BGA feature fusion network was adopted to enhance the BLS-FV-FB model, resulting in a smaller but significant improvement in performance: mAP@0.5 increased by 0.4% to 96.70%, while mAP@0.5(0.5-0.95) improved from 73.80% to 74.60%. This demonstrates that FPN-BGA not only significantly enhances the multi-scale feature information fusion capability but also improves the detection sensitivity for various steel defects, thereby reducing missed detections and improving detection accuracy. This systematic ablation study quantitatively verifies the cumulative benefits of our proposed architecture optimization scheme. The progressive performance enhancements across metrics provide empirical evidence for the effectiveness of each component: optimized detection head design, advanced feature extraction module, and improved multi-scale fusion network.

[0180] Table 4

[0181]

[0182] This invention focuses on the core problems of traditional algorithms in small target detection scenarios of intelligent equipment in coal mines, such as large model parameter volume, slow inference, high interaction delay, and high false detection and false negative rates. In view of the complex environment characteristics of coal mines, such as uneven lighting, dust, and limited field of view, a small target detection algorithm based on SPPF-Vannet and FPN-BGA (SPPF-Vannet-BGA) is proposed. Through multi-dimensional technical innovation and optimization, the environmental noise suppression capability and detection accuracy are improved, and a lightweight network design mechanism is adopted to take into account the requirements of real-time performance and lightweight design. The framework solves the limitations of traditional methods through three major architectural innovations: (1) Significantly improved feature extraction capability. The SPPF-Vannet network achieves a breakthrough in the feature extraction stage, which can more efficiently represent complex, small and unevenly distributed features. (2) Efficient feature fusion. Through the design of the new multi-scale feature fusion network of FPN-BGA with CSPGS module and Contmix, the ability to extract complex and uneven features and capture global features can be improved, effectively reducing network parameters and floating-point calculation volume. (3) Optimized model head module. The structure of the Detect Head module of the model was optimized, and by introducing SIoU as the regression loss function, the regression accuracy and convergence speed were effectively improved.

[0183] After conducting extensive comparative and ablation experiments, it was fully demonstrated that the SPPF-Vannet-BGA algorithm can be successfully applied to the operational status monitoring of conveyor belt recovery equipment in underground coal mines, realizing the detection of abnormal operational status and personnel intrusion, and providing a reliable visual perception solution for the safe operation of intelligent underground equipment. In the future, it can be further extended to multimodal data fusion detection in more scenarios such as underground inspection robots and tunneling equipment. The lightweight design of the SPPF-Vannet-BGA algorithm helps to optimize low-computing-power edge deployment, adapting to more complex and extreme underground environments, and contributing to the improvement of intelligent mining levels.

Claims

1. A target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA, characterized in that: It includes four modules: an image preprocessing module, a feature extraction network module based on SPPF-Vannet, a small target feature fusion and enhancement network module based on CSPGS, ContMix, BIFPN and CA attention mechanism, and a target classification and bounding box prediction module. The image preprocessing module is used to preprocess the acquired raw images containing the target. It combines the illumination estimation module (IEM) and the Mosaic data augmentation method to process the raw images. It stitches together a set of images by random scaling, cropping and arranging, thereby combining multiple images. This expands the dataset while improving the training speed of the model and reducing memory requirements. The feature extraction network module based on SPPF-Vannet: On the basis of the VanillaNet module, a fast spatial pyramid pooling module integrating the SEAM module is fused to construct a feature extraction network module. The preprocessed image is input into this module to suppress the interference of background noise in small target images and extract features with high representation ability. Then, the features from different layers are input into the feature enhancement and fusion module. The small target feature fusion enhancement network module based on CSPGS, ContMix, BIFPN and CA attention mechanism: It includes 4 CSPGS modules, 4 BiFPN convolutional modules, 2 ContMix modules, 1 CA attention mechanism module and 2 Upsample modules; it is used to improve the ability to extract complex non-uniform features and capture global features, and can effectively reduce network parameters and floating-point computation. The target classification and bounding box prediction module consists of three target prediction branches based on the outputs of the three CSPCGS modules of the feature fusion enhancement network module. Different anchor boxes are used to predict targets of different sizes. Each anchor box corresponds to a target of a specific size. During training, in order to accelerate the convergence speed of the target detection model and improve the regression accuracy of the target prediction anchor boxes, the SIOU loss function is minimized to optimize and improve the detection accuracy.

2. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 1, characterized in that: The image preprocessing module specifically includes: By adding an illumination estimation module (IEM), the input image is enhanced in low light conditions. The enhanced image is then augmented using the Mosaic data augmentation method, which stitches together a set of images by randomly scaling, cropping, and arranging them to combine multiple images. In computer vision, image generation can be simplified to a formula similar to Retinex theory: S = R⊙L, where S is the input image, R is the intrinsic reflectivity of the object, and L is the illumination component. The goal of Image Image Prediction (IEM) is to estimate the illumination L from a noisy S to achieve image brightness restoration and detail enhancement. IEM employs an iterative optimization approach, with input u... k via hot start function "Obtain initial illumination" in, Input CNN module C IEM Calculate the regularized gradient ";pass Optimized lighting t k+1 ; use t k+1 Perform pixel-by-pixel division with the original input y to obtain the intermediate image u for the next round. k+1 Repeat the above process until, after K iterations, the final illumination t is output. K and intermediate image u K .

3. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 1, characterized in that: The feature extraction network module based on SPPF-Vannet consists of 5 VanillaNet modules and 1 SPPF-SE module; The network structure was enhanced by introducing a dedicated feature optimization module; the receptive field enhancement module SPPF-SE module was constructed by concatenating the SPPF module with the VanilaNet module and integrating the SEAM attention module into the SPPF module. The SEAM attention module employs depthwise separable convolutions with residual connections. By calculating each channel, it reduces the number of parameters and improves feature extraction capabilities. It fuses the outputs of convolutions at different depths using pointwise convolutions, and then integrates channel features through two fully connected layers to achieve information interaction between channels. It uses an exponential function to process the output of the fully connected layers, expanding the output range of the fully connected layers from [0, 1] to [1, e], so that the effective features of occluded and unoccluded targets are correlated. The attention weights generated by the SEAM attention module are multiplied by the original features. Based on the Vanillanet and SPPF modules, SPPF-Vannet is constructed. This structure is used to extract features from images, enabling the integration of channel, spatial, and location information into the extraction of small target features. This better suppresses background noise interference and enhances the expressive power of the extracted features, providing highly expressive features for subsequent feature enhancement and fusion tasks.

4. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 1, characterized in that: The feature fusion network module FPN-BGA, based on the CSPGS module, ContMix module, BiFPN module and CA attention mechanism, is used to achieve more ideal feature fusion and model lightweighting at multiple scales. It is built upon CSPGS, upsampling, Concat, CA attention mechanisms, and ContMix; its structure includes two branches: The first branch contains two CSPGS modules, two upsampling modules, and two Concat modules. The Concat module is derived from BiFPN and uses learnable normalized weights to perform weighted summation on the input feature maps, thereby achieving the concatenation of feature maps FMs at different scales. The second branch contains 2 CSPGS, 2 ContMix, 2 Concat modules and 1 CA module; By introducing the CA attention mechanism in the first layer of branch 2, the channel weights are dynamically adjusted to highlight key features, thereby enhancing the model's feature discrimination ability and improving the target localization accuracy. At the same time, the ContMix module introduced in branch 2 fully utilizes the contextual priors provided by the overview network, thereby enhancing the ability to comprehensively capture local contextual information. Based on these two branches, the outputs of the three CSPGS modules, namely CSPGS1, CSPGS2 and CSPGS3, will be input into SHead for defect localization and classification. By combining GSConv and ContMix in the multi-scale feature fusion network, features from different levels are effectively aggregated to improve detection performance, significantly improve the detection sensitivity of intelligent equipment targets, and reduce computational costs.

5. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 4, characterized in that: The CSPGS module designs cross-level partial networks using a one-time aggregation method, reducing computational and network structure complexity while maintaining sufficient accuracy. It includes the Conv module and the GSConv module. Depthwise separable convolution (DSC) is based on conventional convolutional neural networks. It divides the computation into two parts: one part is a depthwise convolution applied to a single input channel, and the other part is a regular pointwise convolution with a kernel size of 1x1. The depthwise convolution is performed in a two-dimensional plane, and the number of convolution kernels is the same as the number of channels in the previous layer. The number of feature maps after Depthwise channel-wise convolution is the same as the number of channels in the input layer, which cannot expand the feature maps; Convolution is performed independently on each channel of the input layer, which does not effectively utilize the feature information of different channels at the same spatial location; Pointwise convolution is needed to combine these feature maps to generate new feature maps. Pointwise convolution is very similar to regular convolution. Its kernel size is 1×1×M, where M is the number of channels in the previous layer. Therefore, the convolution operation here will weight and combine the feature maps from the previous step in the depth direction to generate new feature maps. The number of output feature maps is equal to the number of convolution kernels.

6. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 4, characterized in that: To enable convolutions to adapt to different input resolutions and preserve strong local inductive biases, and to fully utilize the contextual priors provided by the overview network, a context-mixed dynamic convolution, ContMix, is introduced. This approach represents the connection between each token in the feature map and the center tokens of multiple regions by calculating an affinity map. This affinity map is then converted into a dynamic convolutional kernel in a learnable manner, and global contextual information is injected into each weight within the kernel. When the dynamic convolutional kernel operates on the feature map through a sliding window, each token is modulated with global information. Furthermore, when such dynamic kernels operate on the feature map through a sliding window, each element in the feature map is modulated by approximate global information collected through the region centers, thus effectively modeling long-range dependencies.

7. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 4, characterized in that: Bidirectional Feature Pyramid Network (BiFPN): BiFPN removes nodes with only one in-degree, eliminating some redundant computations; it adds some residual connections, where original feature nodes in the same layer are connected to the output node to participate in bottom-up feature fusion, and top-down and bottom-up fusion are constructed into a module, which is repeatedly stacked to enhance information fusion; deep features are continuously upsampled and fused with shallow features, and after upsampling, shallow features are continuously downsampled and fused with deep features; in addition, the network adds two lateral connection paths on the basis of the original FPN+PAN structure to integrate the features extracted by the backbone network into the feature map to be detected.

8. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 4, characterized in that: The Coordinate Attention (CA) mechanism decomposes channel attention into a feature encoding process that runs in parallel along two directions: horizontal and vertical. This process establishes long-range dependencies in one direction while preserving positional information in the other. The two are then concatenated to obtain an attention map that is both direction-aware and positional. Finally, after being processed by BN and Sigmoid functions, the attention map is applied to the input feature map to enhance the expressive power of the features. Given an input X, each channel is encoded along the horizontal and vertical directions using pooling kernels of size (H,1) and (1,W), respectively. The output of the c-th channel at width w and height h can be expressed by the following formula: The two encoded feature maps are concatenated and then transformed using a 1×1 convolution function F1, as shown in the following formula: f=δ(F1([z h ,z w ])) (4) In the formula, f represents the feature mapping result of spatial information in the horizontal and vertical directions, δ refers to the nonlinear activation function, and [·,·] is the feature aggregation in space; f decomposition produces f h ∈R C / r×H and f w ∈R C / r×W Two independent features; Using two 1×1 convolution functions F w and F h f respectively w and f h The number of channels is transformed to match the number of channels in the input X, as shown in formula (3): In the formula, σ is the Sigmoid function; g w and g h The expanded output formula of the CA module is as follows:

9. The target detection method for intelligent equipment in coal mines based on SPPF-Vannet and FPN-BGA according to claim 1, characterized in that: The expression for minimizing the SIOU loss function is as follows: Where IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box, Δ is the distance cost, and Ω is the shape cost.

Images