Intelligent mine target detection method and system based on simulation enhancement

Abstract

The invention relates to the technical field of target detection, in particular to an intelligent mine target detection method and system based on simulation enhancement. According to the method, the real mine image is acquired, the simulation image which is difficult to shoot actually is generated based on the real mine image, the simulation image is enhanced and then mixed with the real image to construct the data set, the sample diversity of the data set is expanded, and then the target detection model constructed by the method is trained; the cross-domain generalization ability, weak category recognition precision and model cross-domain robustness of a target detection model can be improved. Meanwhile, a target detection model is constructed on the basis of YOLOv11, through model improvement and optimization of a data set used for model training, the detection capacity of the model for small targets and targets under a complex background is improved, and low-delay and high-frame-rate real-time detection can be achieved. According to the method, simulation enhancement and lightweight model optimization are combined, so that the detection precision, the cross-domain robustness and the real-time performance in a complex mine environment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and more specifically, to a simulation-enhanced smart mine target detection method and system. Background Technology

[0002] With the continuous advancement of intelligent technologies, open-pit mines are accelerating their transformation towards unmanned operation, with driverless mining trucks and autonomous engineering machinery gradually being put into operation on-site. However, open-pit mine roads are unstructured, lacking fixed markings, and often plagued by potholes, falling rocks, and other temporary obstacles. Furthermore, pedestrians, auxiliary vehicles, and large equipment frequently intersect within the work area. These targets vary greatly in size and shape, and are accompanied by complex environmental factors such as dust and low light levels, making them highly susceptible to collisions, vehicle entrapment, and other safety accidents, resulting in equipment damage and personnel injuries. Therefore, achieving high-precision, real-time obstacle detection in unstructured road scenarios has become a crucial link in ensuring safe mine operations and production efficiency.

[0003] Currently, the smart mining field commonly employs multi-sensor fusion solutions combining visible light / infrared cameras, millimeter-wave radar, and lidar, with deep learning vision algorithms serving as the core perception module. In particular, the YOLO (You Only LookOnce) series of single-stage detection frameworks is widely used due to its simple structure and fast detection speed. However, these algorithms are mostly trained based on urban traffic or structured industrial scenarios, lacking specific optimization for open-pit mines. Therefore, their application in mining areas often suffers from poor environmental adaptability, insufficient detection accuracy, and low deployment efficiency. For example, in typical mining conditions such as low light at night, dust obstruction, and distant small targets, the models are prone to false positives and false negatives, failing to meet actual deployment requirements.

[0004] Meanwhile, real-world mine data collection suffers from high costs, uneven sample distribution, and difficulty in reproducing hazardous scenarios, making it difficult for model training to cover all working conditions. Existing methods mainly focus on improving network structure, paying insufficient attention to the scalability and diversity of the data source itself, making it difficult to balance accuracy and real-time performance on embedded devices.

[0005] Therefore, there is an urgent need for a smart mine target detection method that can enhance sample diversity and improve model cross-domain robustness through simulation, and can be efficiently deployed on edge devices, in order to solve the problem of difficulty in balancing detection accuracy and real-time performance under complex open-pit conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a simulation-enhanced intelligent mine target detection method and system to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a simulation-enhanced smart mine target detection method, including: Acquire a first image set that includes multiple images of actual mining scenes; Simulation is performed based on the first image set to obtain a second image set including multiple simulated images; Each image in the third image set is labeled with a target to obtain the third image set with target labeling. The third image set includes the first image set and the second image set. The images in the second image set after target annotation are enhanced using simulation to obtain the fourth image set; A dataset is constructed based on the second image set with target annotations and the fourth image set, thus obtaining the dataset. The preset target detection model is trained based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model. A first image is acquired and input into the trained target detection model to obtain the target category, target category confidence score, and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

[0007] Furthermore, after acquiring the first image set, which includes multiple actual mine scene images, the process further includes: The second image in the first image set is removed; the second image is an image with abnormal resolution, extreme lighting, or severe occlusion. After removing the second image, the images in the first image set are preprocessed, and simulation is performed based on the preprocessed first image set. The preprocessing includes image size unification, format standardization, and color space standardization.

[0008] Furthermore, the simulation based on the first image set includes: The second information is extracted from the first image set to obtain the second information, which includes the illumination direction, color distribution, texture features and target size information of the third image in the first image set; Based on the second information, relevant scene parameters are adjusted in the 3D simulation engine, including ambient lighting, surface material, and camera view. The second image set is generated by the 3D simulation engine after adjusting the relevant scene parameters.

[0009] Furthermore, the target annotation for each image in the third image set includes: By using a preset image annotation tool, each image in the third image set is labeled with a target, and the annotation categories and bounding box coordinates of multiple targets in each image of the third image set are obtained. The annotation categories include heavy engineering equipment, mixed-traffic civilian vehicles, safety workers, and safety warning signs.

[0010] Furthermore, the simulation enhancement of the images in the second image set includes: A preset brightness and contrast adjustment is applied to the fourth image, which is an image from the second image set; Gaussian blur and lens distortion perturbation are added to the fourth image; Add texture noise to the fourth image; Color shift perturbation and style transfer operations are applied to the fourth image.

[0011] Furthermore, the target detection model includes: A feature extraction backbone network is used to extract features from the images in the dataset to obtain multi-layer feature maps. The multi-layer feature maps include low-level detail information, mid-level semantic information, and high-level contextual information of the fifth image, which is an image in the dataset. The feature fusion module is used to perform spatial size alignment on the multi-layer feature map, and to upsample the deep small-size features of the multi-layer feature map and concatenate them with the shallow large-size features of the multi-layer feature map in the channel dimension to obtain a hybrid feature map containing local detail information and high semantic information. A detection head is used to receive the mixed feature map and perform prediction tasks at multiple scales based on the mixed feature map to obtain detection results. The detection results include the target categories, target category confidence scores, and bounding box coordinates of multiple targets in the fifth image. Each detection result corresponds to one fifth image.

[0012] Furthermore, the feature extraction backbone network includes: The initial feature extraction module is used to perform spatial downsampling and expand the number of channels on the fifth image through convolution operations to obtain a first feature map; The feature processing module consists of four layers. Each layer includes a first feature processing module and a second feature processing module. The first feature processing module is used to perform convolutional downsampling on the input feature map to obtain a second feature map. The second feature processing module is used to perform local filtering and information compression on the multi-scale response of the second feature map through separable convolution to obtain a third feature map. The pooling and stitching module is used to receive the fourth feature map, perform multi-scale pooling on the fourth feature map, and stitch the statistical information of different local views in the fourth feature map to the same spatial position in the channel direction. The fourth feature map is the third feature map output by the last layer of the feature processing module.

[0013] Furthermore, training the preset target detection model based on the dataset includes: Multiple fifth image data are input into the target detection model to obtain multiple detection results. Based on the multiple fifth images and the detection results corresponding to the fifth images, and a preset loss function, the target detection model is optimized until the loss function converges, thus obtaining the trained target detection model.

[0014] Furthermore, the loss function is expressed as: ; in, The value of the loss function. The intersection-over-union ratio (IoU) represents the ratio of the predicted bounding boxes to the ground truth bounding boxes. The predicted bounding boxes are the coordinates of the bounding boxes of multiple targets in the detection result corresponding to the fifth image, and the ground truth bounding boxes are the coordinates of the bounding boxes of multiple targets in the fifth image after target annotation. It is the intersection-union ratio that incorporates distance weighting factors. The evaluation criteria are used to assign higher positioning weights when the target center is close; The localization loss is used to measure the similarity between the predicted anchor frame target and the actual anchor frame. It is a location loss The exponential moving average is used to dynamically adjust the focusing mechanism; It is a hyperparameter used to control the intensity of focusing.

[0015] Secondly, this application provides a simulation-enhanced smart mine target detection system, characterized in that it includes: The first acquisition module is used to acquire a first image set including multiple actual mine scene images; The first processing module is used to perform simulation based on the first image set to obtain a second image set including multiple simulated images; The second processing module is used to perform target annotation on each image of the third image set to obtain the target-annotated third image set, which includes the first image set and the second image set. The third processing module is used to perform simulation enhancement on the images of the second image set to obtain the fourth image set; The fourth processing module is used to construct a dataset based on the second atlas and the fourth image set to obtain the dataset; The fifth processing module is used to train a preset target detection model based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model. The sixth processing module is used to acquire the first image and input the first image into the trained target detection model to obtain the target category, target category confidence and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

[0016] The beneficial effects of this invention are as follows: This invention acquires real mine images and generates simulated images that are difficult to capture in real-world conditions based on these images. The simulated images are then enhanced and mixed with the real images to construct a dataset, expanding the sample diversity of the dataset. This dataset is then used to train the target detection model constructed in this invention, improving its cross-domain generalization ability, weak class recognition accuracy, and cross-domain robustness. Furthermore, this invention is based on YOLOv11 and improves the model's backbone feature extraction network and feature fusion module to construct the target detection model. Training this model with the mixed dataset of real and simulated images enhances its ability to detect small targets and targets in complex backgrounds, achieving low-latency, high-frame-rate real-time detection. This invention, through the combination of simulation enhancement and lightweight model optimization, effectively improves detection accuracy, cross-domain robustness, and real-time performance in complex mining environments.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Figure 2 This is a schematic diagram of the YOLOv11-GS model of the simulation-enhanced smart mine target detection method described in this embodiment of the invention; Figure 3This is a schematic diagram of the RepGSNet backbone network of the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Figure 4 This is a schematic diagram of the RepDWConv module of the simulation-enhanced smart mine target detection method described in this embodiment of the invention; Figure 5 This is a schematic diagram of the network structure of the GSNeXt module in the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Figure 6 This is a schematic diagram of the equivalent structure of the multi-branch dilated convolution inference stage of the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Figure 7 This is a schematic diagram showing the model training performance of the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Figure 8 This diagram illustrates a comparison between the improved YOLOv11 model and the training result metrics of YOLOv11 in the simulation-enhanced smart mine target detection method described in this embodiment of the invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] Example 1: This embodiment provides a simulation-enhanced smart mine target detection method.

[0023] See Figure 1The figure shows that the method includes steps S1, S2, S3, S4, S5, S6 and S7.

[0024] S1. Obtain a first image set including multiple actual mine scene images; Specifically, the source of the actual images in this embodiment can be local acquisition, publicly available mining datasets, or historical video frames within the enterprise / platform.

[0025] Furthermore, undesirable images with problems such as abnormal resolution, extreme lighting, or severe occlusion in the acquired actual mine scene image data are removed. The remaining image data undergoes operations such as image size unification, format standardization, and color space standardization to ensure data consistency between subsequent simulation generation and model training.

[0026] Understandably, after acquiring the actual images, they need to be screened and preprocessed to remove samples with abnormal resolution, extreme lighting, or severe occlusion to ensure data quality. Preferably, this embodiment utilizes a custom UnrealCV plugin and a batch control script written in Python to achieve simulated camera automatic navigation and multi-angle shooting. The exported image data is uniformly formatted as JPEG or PNG, with a size of 1920. 1200, to ensure compatibility and input consistency.

[0027] S2. Perform simulation based on the first image set to obtain a second image set including multiple simulated images; Specifically, in this embodiment, step S2 includes: S21. Extract second information from the first image set to obtain the second information, which includes the illumination direction, color distribution, texture features and target size information of the third image in the first image set; S22. Based on the second information, adjust the relevant scene parameters in the 3D simulation engine. The relevant scene parameters include ambient lighting, surface material, and camera view. S23. The second image set is generated by the 3D simulation engine after adjusting the relevant scene parameters, thus obtaining the second image set.

[0028] Specifically, in this embodiment, after screening and preprocessing the actual mine scene images, based on the characteristics of the actual mining images, simulated mine target image data based on a 3D simulation engine is generated. The simulated images are generated by Unreal Engine 5 (UE5), a 3D simulation engine. By controlling scene elements (terrain, equipment models, physical materials), camera parameters (viewpoint, focal length, pixel resolution), and environmental variables (dust, haze, lighting), simulated image data containing typical mining conditions (nighttime, dust, occlusion, distant small targets, etc.) is generated. The simulation data is mainly used to supplement scene samples that are difficult to collect or reproduce in the real environment, such as dangerous working conditions (special operations, nighttime construction), rare perspectives (side and rear, extreme distance), and weakly classified targets (workers, warning signs, civilian vehicles), thereby significantly expanding the multi-scene coverage and category balance of the dataset.

[0029] Understandably, based on the illumination direction, color distribution, texture features, and target size information extracted from actual mining images, relevant scene parameters, including ambient lighting, surface material, and camera viewpoint, are adjusted in the 3D simulation engine to ensure that the simulated scene matches the real mining area images in terms of visual statistical features. This generates simulated mining target image data across multiple weather conditions, time periods, and angles. Furthermore, by performing intra-domain mapping of actual mining feature parameters in the 3D simulation engine, controllable combinations of shooting conditions, target distribution, and ambient lighting are achieved. Unlike traditional random augmentation, this method uses the statistical characteristics of real samples as constraints, ensuring that the simulated samples maintain consistency with real mining area data in terms of brightness distribution, texture structure, and target scale. This allows for the reconstruction of hazardous working conditions, rare viewpoints, and weakly classified target samples in the real environment within the simulation space, achieving data-level completion and balance.

[0030] S3. Perform target annotation on each image in the third image set to obtain the third image set after target annotation. The third image set includes the first image set and the second image set. Specifically, in this embodiment, step S3 includes: By using a preset image annotation tool, each image in the third image set is labeled with a target, and the annotation categories and bounding box coordinates of multiple targets in each image of the third image set are obtained. The annotation categories include heavy engineering equipment, mixed-traffic civilian vehicles, safety workers, and safety warning signs.

[0031] Understandably, in this embodiment, the collected and simulated data are filtered and uniformly labeled. The LabelImg open-source tool is used to annotate all images, thereby obtaining the target categories and corresponding bounding box coordinates of multiple targets in the images, which are then used for training the target detection model. A YOLO-formatted annotation information txt file is output, with annotation categories including: heavy engineering equipment (mining trucks, excavators, bulldozers, and trailers), mixed-use civilian vehicles, safety personnel, and safety warning signs—targets of interest in the actual mining scene.

[0032] S4. Perform simulation enhancement on the images of the second image set after target annotation to obtain the fourth image set; Specifically, in this embodiment, step S4 includes: Apply preset brightness and contrast adjustments to the second image data; Gaussian blur and lens distortion perturbation are added to the second image data; Add texture noise to the second image data; Color shift perturbation and style transfer operations are applied to the second image data.

[0033] It is understandable that enhancement processing is applied to simulated images, including brightness / contrast adjustment, color cast and white balance perturbation, random noise superposition, Gaussian blur, lens distortion, occlusion superposition, and style transfer, to improve the diversity of simulated data and its domain consistency with real data. Preferably, in this embodiment, the simulated images undergo targeted enhancement processing. Unlike general image enhancement, this scheme, based on the characteristics of high contrast, low noise, and overly idealized textures exhibited by the mine simulation samples during rendering, proposes four steps: illumination perturbation, imaging degradation, interference superposition, and style unification. This makes the simulated images more closely resemble real samples in terms of visual domain distribution, thereby significantly improving the model's cross-domain detection performance.

[0034] Illumination perturbation: Apply random adjustments to the brightness and contrast of the simulated image to simulate the environmental changes of the mining area, such as day and night alternation, strong backlight, and uneven illumination. Image degradation: Gaussian blur and lens distortion perturbation are added to the simulated image to reproduce the image degradation effect of a real camera and enhance the uncertainty of image edges; Interference overlay: Add texture noise such as dust and occlusion to the simulated image to simulate the visibility interference and target occlusion common in mining areas; Style unification: Color shift perturbation and style transfer operations are applied to the simulated images to reduce domain bias; specifically, a unified style processing strategy is introduced. Since there is domain bias between real and simulated images, this invention achieves cross-domain fusion through image enhancement. In the image space, tone normalization is uniformly used to preprocess both real and simulated data, which can be expressed as: ; in, For simulated images, and Real images The mean and standard deviation, and For simulated images The statistics.

[0035] S5. Construct a dataset based on the second image set after target annotation and the fourth image set to obtain the dataset; Specifically, in this embodiment, step S5 includes: Enhanced simulated images and real-world images are mixed in a preset ratio to construct training and validation datasets. Preferably, simulated data accounts for approximately 20%, and real data accounts for approximately 80%. This supplements complex working condition samples while maintaining the dominance of real samples, improving the model's cross-domain generalization ability and weak class recognition accuracy. The combination of the environmental diversity of real images and the label accuracy of simulated images overcomes the problems of scarce data, insufficient weak class samples, and difficulty in collecting samples from dangerous scenes in traditional mining areas. This significantly improves the target detection model's ability to detect small targets, its cross-domain generalization performance, and its safety scene recognition coverage under complex open-pit conditions. During training, random sampling is used to input the training process to avoid training bias. Uniform preprocessing can be used before and after mixing to further reduce domain differences and improve cross-domain robustness.

[0036] S6. Train the preset target detection model based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model; It is worth noting that, in this embodiment, as Figure 2 As shown, the object detection model YOLOv11-GS includes: A feature extraction backbone network is used to extract features from the images in the dataset to obtain multi-layer feature maps. The multi-layer feature maps include low-level detail information, mid-level semantic information, and high-level contextual information of the fifth image, which is an image in the dataset. The feature fusion module is used to perform spatial size alignment on the multi-layer feature map, and to upsample the deep small-size features of the multi-layer feature map and concatenate them with the shallow large-size features of the multi-layer feature map in the channel dimension to obtain a hybrid feature map containing local detail information and high semantic information. A detection head is used to receive the mixed feature map and perform prediction tasks at multiple scales based on the mixed feature map to obtain detection results. The detection results include the target categories, target category confidence scores, and bounding box coordinates of multiple targets in the fifth image. Each detection result corresponds to one fifth image.

[0037] Furthermore, the feature extraction backbone network includes: The initial feature extraction module is used to perform spatial downsampling and expand the number of channels on the fifth image through convolution operations to obtain a first feature map; The feature processing module consists of four layers. Each layer includes a first feature processing module and a second feature processing module. The first feature processing module is used to perform convolutional downsampling on the input feature map to obtain a second feature map. The second feature processing module is used to perform local filtering and information compression on the multi-scale response of the second feature map through separable convolution to obtain a third feature map. The pooling and stitching module is used to receive the fourth feature map, perform multi-scale pooling on the fourth feature map, and stitch the statistical information of different local views in the fourth feature map to the same spatial position in the channel direction. The fourth feature map is the third feature map output by the last layer of the feature processing module.

[0038] Specifically, the target detection model in this embodiment improves the traditional feature extraction module's feature extraction backbone network, naming it the RepGSNet backbone network. It performs multi-scale feature extraction on images in the dataset, and the multi-layer feature map includes low-level detail information, mid-level semantic information, and high-level contextual information of the image.

[0039] Furthermore, the target detection model of this embodiment will be described. After the dataset is constructed, the preprocessed and mixed real images and simulated images are input into the YOLOv11-GS model of this embodiment.

[0040] First, the input image is fed into the RepGSNet backbone network. For example... Figure 3 As shown, the backbone network is used to extract multi-level feature representations. The overall structure consists of an initial Stem encoding layer, four feature extraction stages (Stage 1 to Stage 4), and a final SPPF module. Wherein: The input image is received by the Stem module and subjected to preliminary convolutional encoding to extract low-level texture and edge features and normalize the input channels.

[0041] Subsequently, the features enter four stages, from Stage 1 to Stage 4, to complete feature downsampling and semantic abstraction in sequence. Each stage consists of a downsampling module and alternating stacked GSFBlock and GSEBlock.

[0042] The downsampling module reduces the spatial resolution of the feature map through convolution operations, achieving scale compression and semantic enhancement. GSFBlock Lightweight Multi-Scale Convolution Module: Expands the receptive field and improves the ability to represent local features while maintaining computational efficiency through multi-scale reparameterized convolutional structures; GSEBlock Large Kernel Channel Attention Module: Introduces a channel attention mechanism (Squeeze-and-Excitation) to adaptively adjust channel weights, enhance key semantic feature responses, and suppress redundant information.

[0043] The multi-layer feature maps extracted by the RepGSNet backbone network contain low-level detail information, mid-level semantic information, and high-level contextual information. These features are then fed into the feature fusion enhancement neck.

[0044] Understandably, this embodiment uses an improved YOLOv11-GS network structure to build an object detection model. RepGSNet is used as the backbone feature extraction network, internally employing alternating stacks of GSEBlock and GSFBlock to achieve deep semantic mining and channel attention enhancement. Specifically, GSFBlock embeds dilated reparameterized convolutions, while GSEBlock introduces an SE channel attention module. A lightweight fusion module, GSNeXt, is introduced at the model's neck, combining large-kernel deep convolutions with inverted bottleneck pointwise convolutions to achieve multi-scale feature aggregation and contextual modeling, thereby enhancing the detection performance of small and weakly classified targets in complex mining scenes.

[0045] Specifically, after the dataset is constructed, the preprocessed and mixed real and simulated images are input into the improved YOLOv11-GS model. The backbone uses the proposed RepGSNet to extract robust structural and semantic information in complex multi-scale scenes. The input fourth image data first enters the Stem initial feature extraction module. Stem performs spatial downsampling and expands the number of channels of the image through convolution operations, thereby improving the original image quality. Input converted The first feature map. This feature map, while preserving key textures and edge structures, removes some high-frequency noise, providing a stable low-level visual representation for subsequent deep feature learning.

[0046] Subsequently, the first feature map is sequentially processed through Stages 1 through 4 for hierarchical feature extraction. Taking any Stage as an example, this stage first performs convolutional downsampling on the input feature map, reducing the spatial size from... Reduced to The reduced spatial density of the feature map, coupled with a significantly expanded receptive field, integrates previously scattered local gradients into a larger statistical response, resulting in enhanced spatial correlation and a shift of local variance to a larger scale in the data distribution. The downsampled features are then processed using alternating stacked GSFBlock and GSEBlock algorithms. Compared to the input, the processed output feature map exhibits increased inter-channel mutual information and a higher rank in the covariance matrix, reflecting the superposition of multi-scale semantics.

[0047] GSFBlock expands the receptive field on the same feature map by sequentially stacking 3×3, 5×5, and 7×7 dilated convolutions, thereby continuously mapping the original local response into a multi-scale structural response in the spatial dimension. At the data level, GSFBlock's processing transforms the spatial response distribution within a single channel from "primarily focused on local details" to a hybrid distribution "containing both global contours and local details." While the explicit expansion of multi-scale information may temporarily increase redundancy between channels, it provides richer contextual information for subsequent discriminative reallocation. The intention of this step is to achieve a receptive field expansion effect comparable to multi-branch operations with single-branch sequential operations, while avoiding the size alignment and channel matching problems common in multi-branch fusion, thus maintaining inference efficiency and module connectability in engineering implementation.

[0048] GSEBlock first uses depthwise separable convolution (DWConv) to locally filter and compress the multi-scale response after GSEBlock expansion. This moderately smooths and reshapes the long-range correlation caused by the large receptive field in space, statistically reducing high-frequency noise components and improving the local signal-to-noise ratio of significant structures. Subsequently, the SE channel attention mechanism adaptively weights each channel based on the global channel description vector, amplifying the discriminative channel response while suppressing low-value or noise-dominated channels. The result is represented in the data representation as sparsity of channel responses, decrease in channel entropy, and redistribution of mutual information between channels. Through this sequential processing flow of "space → local reshaping → channel recalibration," the output features simultaneously improve discriminativeness and noise resistance while maintaining multi-scale structural information.

[0049] After the four stages are completed, the deep features from Stage 4 enter the SPPF module. SPPF uses multi-scale pooling and concatenation in the channel direction to juxtapose statistical information from different local views to the same spatial location, so that single-point features contain statistical summaries from different scale contexts. This expands the context coverage of each pixel without changing the spatial resolution, which is convenient for subsequent necking to accurately align and fuse cross-scale semantics.

[0050] Finally, the shallow, mid-level, and deep features output by the RepGSNet backbone are multi-scale aligned and fused at the GSNeXt neck. GSNeXt uses upsampling, spatial alignment, and channel concatenation, combined with lightweight multi-scale convolution, to readjust the distribution of features at each scale, making the semantics at different resolutions numerically consistent. This facilitates a robust input distribution for the detector head in classification and regression tasks. After this series of data processing and re-encoding, the features received by the detector head are characterized by noise suppression, enhanced structural information, and a more sparse and concentrated semantic distribution, thus exhibiting higher discriminative power in object classification, bounding box regression, and confidence estimation.

[0051] Furthermore, it can be understood that the improved feature extraction backbone network RepGSNet in this embodiment combines multi-scale reparameterized convolution and SE channel attention mechanism, achieving a larger effective receptive field, stronger feature representation ability, and higher parameter efficiency through structural optimization. The backbone structure is divided into four stages (Stage 1 to Stage 4), each stage reducing spatial dimensionality and improving semantic abstraction ability through downsampling operations. Within each stage, this invention uses GSFBlock and GSEBlock modules stacked sequentially to form feature units, creating a progressive feature extraction path from spatial multi-scale enhancement to channel semantic enhancement, ensuring modeling capability while effectively reducing computational cost.

[0052] The overall network data flow path is as follows: X→Stem→Stage1→Downsample1→Stage2→···→Stage4.

[0053] It is worth noting that the GSEBlock module further optimizes the multi-scale fused feature map through a channel attention mechanism, focusing on strengthening the response to key semantic channels and suppressing noise features caused by redundant background and illumination perturbations. Open-pit mine environments are highly susceptible to background interference. In actual working conditions, frequent changes in illumination (such as strong and weak light during the day and night, front and back lighting), occlusion (such as dust and vehicles), and differences in obstacle materials all contribute to the instability of feature representation. In open-pit mine environments with drastic lighting changes, frequent dust occlusion, and complex background textures, traditional deep convolutional features are often easily affected. To adapt to the complex and variable working environment of open-pit mines, this invention designs a multi-scale feature extraction unit composed of GSFBlock and GSEBlock stacked sequentially, which exhibits stronger robustness in such scenarios.

[0054] Furthermore, the specific descriptions of the GSFBlock and GSEBlock modules are as follows: GSFBlock expands the effective receptive field of features through multi-scale convolutional kernels and dilated convolutional groups, enabling it to model target contours, shadow boundaries, and complex structural textures. All convolutions use stride=1, and padding is set according to the kernel size to ensure that the output maintains a consistent spatial size with the input. The output of GSFBlock is fed as input into the subsequent GSEBlock.

[0055] Unlike GSFBlock's spatial modeling, GSEBlock aims to enhance channel semantics. Internally, it consists of depthwise convolutions and SE channel attention.

[0056] Specifically, for the input feature map GSEBlock first obtains the following through depthwise convolution: ; Among them, DWConv is a depthwise separable convolution, which has a good ability to capture local textures, edge contours, and dust-occluded broken shapes in mining scenes, while having extremely low overhead, making it suitable for embedded deployment.

[0057] Subsequently, the importance of the channel dimensions is redistributed using the SE module. The Squeeze operation is applied to the depthwise convolution output. Perform global average pooling to obtain the channel description vector. , can be represented as: ; in, Indicates the first The output values ​​of each channel after global average pooling, where H represents the height of the feature map, W represents the width of the feature map, and H×W represents the number of spatial locations in the feature map. Represents the first element in the input feature map. Each channel is located in eigenvalues; Next, the excitation operation constructs a nonlinear channel mapping using two fully connected layers, and the channel weight vector is generated by the sigmoid function. , can be represented as: ; in, Represents a dimension reduction matrix. This represents the upscaling matrix, where r is the channel compression ratio; The resulting channel weights will be applied to the input feature map to achieve channel reweighting, i.e., the output... With the same spatial dimensions as the input, the output can be represented as: ; Where Y represents the new feature map after SE attention filtering, and c is the channel dimension. represents the weighting coefficient corresponding to the c-th channel. This weighting strategy enhances key semantics in the mining area (such as the outline of operating equipment, personnel areas, and significantly moving objects) and suppresses lighting variations, dust noise, and high-texture interference.

[0058] GSEBlock and GSFBlock use a sequential stacking structure: →GSFBlock→GSEBlock→ That is, the input of GSEBlock comes from the output of GSFBlock. Both modules use stride=1 and the same padding to ensure that the output space size is consistent with the input, which is suitable for the resolution requirements of the YOLOv11 backbone network.

[0059] Furthermore, RepGSNet employs a smaller scaling ratio and appropriately increases channel width to reduce channel redundancy in the feedforward network (FFN), thereby improving inference efficiency and actual deployment performance on the device side. For example, the scaling ratio is set to 2 or 3, and a channel doubling strategy is used at the end of each stage to compensate for the loss of expressive power caused by the reduced scaling ratio. This scheme can improve overall detection accuracy while maintaining low latency.

[0060] By combining multi-scale spatial augmentation of GSFBlock with channel-wise semantic selection of GSEBlock, the RepGSNet backbone of this invention can better adapt to extreme scenarios in open-pit mines, such as small targets, weak textures, strong illumination variations, and dust occlusion. This structure achieves synergistic optimization of shallow spatial modeling and deep semantic augmentation, significantly improving the stability and robustness of target detection.

[0061] This involves weighting and enhancing each channel through multiplication. This makes the network more focused on channel features with high discriminative power in open-pit mine scenarios, such as equipment outlines and personnel outlines, while suppressing invalid responses caused by background or lighting, thereby improving the effectiveness of overall feature representation and detection accuracy.

[0062] Furthermore, it is understandable that, such as Figure 4 The RepDWConv module structure diagram shown below is described in detail: GSFBlock employs a multi-scale dilated convolutional fusion structure, which contains multiple parallel depthwise convolutional branches. Each branch corresponds to a different kernel size (e.g., 3×3, 5×5, 7×7) and dilation rate (e.g., d=1, 2, 3). Each branch can extract edge, texture, and global contour features from the image at different spatial scales to capture the spatial diversity of the target. The fusion operation of the above multi-branch convolution can be represented as: ; in, This represents the output feature map after multi-scale convolutional branch fusion. This represents the convolution weights corresponding to different kernel sizes. This represents the convolution weights corresponding to different inflation rates. Indicates the expansion rate The processed input feature map, For bias terms, This indicates a convolution operation.

[0063] GSFBlock significantly expands the effective receptive field by using RepDWConv to dilate convolutions, while keeping the number of parameters lightweight. This allows the network to acquire richer contextual information, thereby enhancing its feature representation capabilities for distant targets, partially occluded targets, and targets with weak textures. It is especially suitable for the detection needs of multi-scale targets such as workers, warning signs, mining trucks, and excavators in open-pit mines.

[0064] In the network structure, RepDWConv constructs a multi-scale convolutional branch of GSFBlock to extract basic spatial structural features of different receptive fields. The output of GSFBlock is directly fed into subsequent GSEBlocks in the same stage, further enhancing channel expressiveness and nonlinear modeling capabilities through depthwise separable convolution, channel attention (SE), and feedforward network (FFN), achieving joint representation of local details and long-range dependencies.

[0065] The multi-scale features of GSFBlock not only determine the semantic modeling basis of subsequent GSEBlock, but also affect the final output quality of each stage, and further affect the multi-scale fusion effect of YOLOv11 neck layer and the performance of the detection head in classification and bounding box regression tasks.

[0066] Furthermore, such as Figure 5 As shown, the feature fusion module GSNeXt performs step-by-step upsampling, size alignment, and cross-scale fusion processing on the multi-layer feature maps from the backbone network to generate structurally consistent multi-scale features for subsequent target classification and bounding box regression.

[0067] Specifically, this embodiment first aligns the feature maps of different scales output by the backbone network in terms of spatial size. By upsampling the deep, small-scale features and concatenating them with the shallow, large-scale features in the channel dimension, a hybrid feature map containing local detail information and high semantic information is obtained. Based on this, a corresponding hybrid feature map is generated for each layer of the multi-layer feature map, and input into the GSNeXt module to complete cross-scale information integration, thereby obtaining a set of structurally consistent and semantically enhanced multi-scale fused features.

[0068] Finally, the multi-scale fusion features are input into the detection head to achieve prediction tasks at multiple scales and obtain corresponding detection results, one of which corresponds to one of the fourth image data.

[0069] It is understandable that, such as Figure 5 As shown, this invention replaces the original ordinary convolution in the neck structure of YOLOv11n with a lightweight multi-scale fusion module GSNeXt, realizing cross-scale fusion processing of feature maps after upsampling. Specifically, for a size of... The Layer feature map , compared with the first after upsampling Layer feature map Channel splicing is performed to obtain a hybrid feature map. Its number of channels is determined by and It is accumulated; then it will be The GSNeXt module is input, and local structural features are extracted through 3×3 depthwise convolutions. Bottleneck-style double 1×1 convolutions are then used for channel expansion and compression, respectively, effectively integrating the spliced ​​features at the data level. The fused output... It not only maintains a uniform spatial size, but also has cross-scale semantic consistency, making it particularly suitable for complex scenes such as low contrast, weak texture, and small-sized targets in open-pit mines.

[0070] Based on the above structure, after the fusion of features at each layer is completed, a multi-scale fused feature set with consistent structure is formed, and finally input into the detection head to complete the classification and bounding box regression tasks.

[0071] Furthermore, it is understandable that, in combination Figures 2 to 5 RepGSNet, as the backbone network in the YOLOv11-GS model, aims to achieve a synergistic unity of local detail extraction and global modeling capabilities through the stacked design of GSFBlock and GSEBlock. This ensures that each stage possesses both spatial scale modeling capabilities and channel semantic focusing capabilities, forming complementary enhancements. The attention-free design of GSFBlock allows it to focus more on parsing the original spatial structure, avoiding premature channel filtering before features have formed stable semantics, thus preserving more basic information.

[0072] GSNeXt, as the neck fusion module in the YOLOv11-GS model, continuously strengthens the supplementation of high-level features by low-level semantic information, which helps to improve the recognition ability of occluded targets and targets in mixed backgrounds. It is an important part of the present invention to achieve efficient detection of small targets in mining areas and low-latency edge deployment.

[0073] The detection head (Detect module) receives three sets of feature maps (P3, P4, P5) from the neck fusion module, which are used to detect small, medium and large targets, respectively.

[0074] The detection output includes: category prediction vector Used for multi-class classification; confidence score , representing the probability of the target's existence; bounding box parameters It is used to locate target bounding boxes. An image will generate multiple prediction boxes, and each prediction box corresponds to a target. Each prediction box has a set of prediction vectors, confidence scores and bounding box parameters.

[0075] Furthermore, in this embodiment, training the target detection model based on the dataset and the first information includes: Multiple fourth image data are input into the target detection model to obtain multiple first detection results. The target detection model is optimized based on the multiple first detection results and a preset loss function until the loss function converges, and the trained target detection model is obtained. The loss function is expressed as: ; in, The value of the loss function. This represents the intersection-over-union ratio (IoU) between the predicted bounding boxes and the ground truth bounding boxes. The predicted bounding boxes are the coordinates of the bounding boxes of multiple targets in the image in the detection results, while the ground truth bounding boxes are the coordinates of the bounding boxes of multiple targets in the image after the targets are labeled. It is the intersection-union ratio that incorporates distance weighting factors. The evaluation criteria are used to assign higher positioning weights when the target center is close; The localization loss is used to measure the similarity between the predicted anchor frame target and the actual anchor frame. It is a location loss The exponential moving average is used to dynamically adjust the focusing mechanism; It is a hyperparameter used to control the intensity of focusing.

[0076] Preferred, such as Figure 6 As shown in this embodiment, the multi-scale dilated convolution module of the model construction includes multiple parallel depthwise convolution branches during the training phase. During the inference phase, based on the principle of equivalent reparameterization of the convolution kernel, the above-mentioned multi-branch convolution kernel is folded into a single 7×7 depthwise separable convolution through operations such as unrolling, zero-padding, and weight stacking. This folded large kernel convolution significantly reduces the complexity of the inference structure and improves deployment efficiency while maintaining the equivalent receptive field during the training phase.

[0077] This folding process does not change the size and channel structure of the feature map, and can be seamlessly integrated with subsequent depthwise convolutions and channel attention modules.

[0078] S7. Acquire a first image and input the first image into the trained target detection model to obtain the target category, target category confidence and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

[0079] Understandably, the trained target detection model can be directly used to automatically detect and identify real-time captured images of mine scenes, continuous video frames, or offline collected images of the work site. During deployment, it can output the corresponding target category and location results based on the input current image data, and can be directly applied to scenarios such as mine operation monitoring, vehicle obstacle avoidance, and risk warning. The model can run on edge computing devices to achieve low-latency real-time on-site detection.

[0080] Furthermore, it is understood that the purpose of this invention is to address the problems of insufficient robustness under complex working conditions, difficulty in sample data acquisition, limited model generalization ability, and poor real-time performance in edge deployment of existing mine target detection methods. To address these issues, this invention proposes a simulation-enhanced smart mine target detection method and system. By introducing high-quality simulation data to expand sample diversity and combining it with targeted network structure improvements, stable detection and identification of dangerous, difficult-to-photograph, and weakly categorized targets can be achieved, thereby improving detection accuracy and robustness in actual deployment scenarios.

[0081] This embodiment tests the performance metrics of the target detection model and compares them with the performance metrics of the base model YOLOv11n, as detailed below: like Figure 7 As shown, during the 0-300 training epochs, the losses in both the training and validation phases gradually decreased from high to low and eventually converged. Simultaneously, the performance metrics increased with each training epoch, stabilizing after approximately 200 epochs, with Precision reaching 0.92 and mAP50 reaching approximately 0.909. The curves in the figure demonstrate that the model training process was stable, with no significant oscillations in the metrics, ultimately achieving a relatively ideal detection accuracy.

[0082] like Figure 8 As shown, comparing the performance metrics of this embodiment's model with the base model YOLOv11n—precision, recall, mAP50, and mAP50-95—it can be seen that during most training phases and the final stable phase, the curves for all four metrics of this embodiment are above those of the comparison model. Therefore, it can be directly concluded that within the same training epochs, this embodiment outperforms the comparison model in terms of precision, recall, and mAP.

[0083] Example 2: Unlike Example 1, Example 2 provides a simulation-enhanced smart mine target detection system, comprising: The first acquisition module is used to acquire a first image set including multiple actual mine scene images; The first processing module is used to perform simulation based on the first image set to obtain a second image set including multiple simulated images; The second processing module is used to perform target annotation on each image of the third image set to obtain the target-annotated third image set, which includes the first image set and the second image set. The third processing module is used to perform simulation enhancement on the images of the second image set to obtain the fourth image set; The fourth processing module is used to construct a dataset based on the second atlas and the fourth image set to obtain the dataset; The fifth processing module is used to train a preset target detection model based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model. The sixth processing module is used to acquire the first image and input the first image into the trained target detection model to obtain the target category, target category confidence and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

[0084] Furthermore, the model is exported and optimized: After training the model in the PyTorch framework, it is exported to the Open Neural Network Exchange Format (ONNX) using the PyTorch framework export interface (torch.onnx.export). Before deployment, the model undergoes optimization processes such as structural pruning, lightweight quantization, operator fusion, and redundant layer removal to ensure compatibility with the TensorRT inference engine and improve inference speed and edge device adaptability.

[0085] Understandably, to verify the effectiveness of the algorithm of this invention, comparative experiments were conducted on the KITTI dataset against mainstream detection models such as YOLO12n, YOLO11n, YOLO11s, YOLOv10n, YOLOv9t, YOLOv8n, YOLOv6n, YOLOv5n, YOLOv3-tiny, Faster R-CNN, and DETR. Experimental results show that, while keeping the number of parameters and computational cost within acceptable ranges, the detection accuracy is significantly improved compared to existing methods. Compared to YOLO12n, the accuracy is improved by approximately 7.9%; compared to YOLO11s, the number of parameters is approximately 30% and the computational cost is approximately 45%, but the accuracy is still improved by approximately 2%; simultaneously, compared to representative two-stage detection models Faster R-CNN and DETR, the algorithm of this invention improves accuracy by approximately 40% and 13.1%, respectively.

[0086] System Deployment and Engineering: This invention utilizes ROS2 to build an edge-aware platform. The system consists of nodes such as image subscription, model inference, result publishing, and visualization. Its deployment process includes the following steps: First, the sensor image is read from the / image_raw topic; the ONNX lightweight model is loaded by calling the ONNXRuntime interface; inference is performed on the Jetson AGX Orin edge device, and the detection results are published to the / detection_result topic to achieve real-time target detection and result distribution.

[0087] By compressing models, exporting from ONNX, and deploying on the ROS2 platform, a complete industrial-grade perception system deployment solution is formed, ensuring that the target detection algorithm can run efficiently and stably on resource-constrained smart mine edge devices, providing solid support for applications such as unmanned operation and intelligent perception.

[0088] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A simulation-enhanced smart mine target detection method, characterized in that, include: Acquire a first image set that includes multiple images of actual mining scenes; Simulation is performed based on the first image set to obtain a second image set including multiple simulated images; Each image in the third image set is labeled with a target to obtain the third image set with target labeling. The third image set includes the first image set and the second image set. The images in the second image set after target annotation are enhanced using simulation to obtain the fourth image set; A dataset is constructed based on the second image set with target annotations and the fourth image set, thus obtaining the dataset. The preset target detection model is trained based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model. A first image is acquired and input into the trained target detection model to obtain the target category, target category confidence, and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

2. The simulation-enhanced smart mine target detection method according to claim 1, characterized in that, After acquiring the first image set, which includes multiple images of actual mine scenes, the process further includes: The second image in the first image set is removed; the second image is an image with abnormal resolution, extreme lighting, or severe occlusion. After removing the second image, the images in the first image set are preprocessed, and simulation is performed based on the preprocessed first image set. The preprocessing includes image size unification, format standardization, and color space standardization.

3. The simulation-enhanced smart mine target detection method according to claim 1, characterized in that, The simulation based on the first image set includes: The second information is extracted from the first image set to obtain the second information, which includes the illumination direction, color distribution, texture features and target size information of the third image in the first image set; Based on the second information, relevant scene parameters are adjusted in the 3D simulation engine, including ambient lighting, surface material, and camera view. The second image set is generated by the 3D simulation engine after adjusting the relevant scene parameters.

4. The simulation-enhanced smart mine target detection method according to claim 1, characterized in that, The target annotation for each image in the third image set includes: By using a preset image annotation tool, each image in the third image set is labeled with a target, and the annotation categories and bounding box coordinates of multiple targets in each image of the third image set are obtained. The annotation categories include heavy engineering equipment, mixed-traffic civilian vehicles, safety workers, and safety warning signs.

5. The simulation-enhanced smart mine target detection method according to claim 1, characterized in that, The simulation enhancement of the images in the second image set includes: A preset brightness and contrast adjustment is applied to the fourth image, which is an image from the second image set; Gaussian blur and lens distortion perturbation are added to the fourth image; Add texture noise to the fourth image; Color shift perturbation and style transfer operations are applied to the fourth image.

6. The simulation-enhanced smart mine target detection method according to claim 1, characterized in that, The target detection model includes: A feature extraction backbone network is used to extract features from the images in the dataset to obtain multi-layer feature maps. The multi-layer feature maps include low-level detail information, mid-level semantic information, and high-level contextual information of the fifth image, which is an image in the dataset. The feature fusion module is used to perform spatial size alignment on the multi-layer feature map, and to upsample the deep small-size features of the multi-layer feature map and concatenate them with the shallow large-size features of the multi-layer feature map in the channel dimension to obtain a hybrid feature map containing local detail information and high semantic information. A detection head is used to receive the mixed feature map and perform prediction tasks at multiple scales based on the mixed feature map to obtain detection results. The detection results include the target categories, target category confidence scores, and bounding box coordinates of multiple targets in the fifth image. Each detection result corresponds to one fifth image.

7. The simulation-enhanced smart mine target detection method according to claim 6, characterized in that, The feature extraction backbone network includes: The initial feature extraction module is used to perform spatial downsampling and expand the number of channels on the fifth image through convolution operations to obtain a first feature map; The feature processing module consists of four layers. Each layer includes a first feature processing module and a second feature processing module. The first feature processing module is used to perform convolutional downsampling on the input feature map to obtain a second feature map. The second feature processing module is used to perform local filtering and information compression on the multi-scale response of the second feature map through separable convolution to obtain a third feature map. The pooling and stitching module is used to receive the fourth feature map, perform multi-scale pooling on the fourth feature map, and stitch the statistical information of different local views in the fourth feature map to the same spatial position in the channel direction. The fourth feature map is the third feature map output by the last layer of the feature processing module.

8. The simulation-enhanced smart mine target detection method according to claim 6, characterized in that, The step of training a preset target detection model based on the dataset includes: Multiple fifth image data are input into the target detection model to obtain multiple detection results. Based on the multiple fifth images and the detection results corresponding to the fifth images, and a preset loss function, the target detection model is optimized until the loss function converges, thus obtaining the trained target detection model.

9. The simulation-enhanced smart mine target detection method according to claim 8, characterized in that, The loss function is expressed as: ； in, The value of the loss function. The intersection-over-union ratio (IoU) represents the ratio of the predicted bounding boxes to the ground truth bounding boxes. The predicted bounding boxes are the coordinates of the bounding boxes of multiple targets in the detection result corresponding to the fifth image, and the ground truth bounding boxes are the coordinates of the bounding boxes of multiple targets in the fifth image after target annotation. It is the intersection-union ratio that incorporates distance weighting factors. The evaluation criteria are used to assign higher positioning weights when the target center is close; The localization loss is used to measure the similarity between the predicted anchor frame target and the actual anchor frame. It is a location loss The exponential moving average is used to dynamically adjust the focusing mechanism; It is a hyperparameter used to control the intensity of focusing.

10. A simulation-enhanced intelligent mine target detection system according to any one of claims 1–9, characterized in that, include: The first acquisition module is used to acquire a first image set including multiple actual mine scene images; The first processing module is used to perform simulation based on the first image set to obtain a second image set including multiple simulated images. The second processing module is used to perform target annotation on each image of the third image set to obtain the target-annotated third image set, which includes the first image set and the second image set. The third processing module is used to perform simulation enhancement on the images of the second image set after target annotation to obtain the fourth image set; The fourth processing module is used to construct a dataset based on the second atlas after target annotation and the fourth image set, thereby obtaining the dataset; The fifth processing module is used to train a preset target detection model based on the dataset to obtain the trained target detection model, which is built based on the YOLOv11 model. The sixth processing module is used to acquire the first image and input the first image into the trained target detection model to obtain the target category, target category confidence and bounding box coordinates of multiple targets in the first image. The first image is a real-time acquired image of an actual mine scene.

Images