A mine conveying belt foreign matter detection method, device, equipment, medium and product

Abstract

This application discloses a method, apparatus, equipment, medium, and product for detecting foreign objects on mine conveyor belts, relating to the fields of computer vision and artificial intelligence. The method includes acquiring original images of the mine conveyor belt area; sequentially performing low-light image enhancement and super-resolution reconstruction processing on the original images to obtain high-quality, high-resolution images; identifying and locating foreign objects in the high-resolution images using an improved target detection model; and outputting detection results including foreign object category labels, confidence scores, and location coordinates. This application achieves high-precision, high-recall real-time detection and accurate location of various foreign objects such as gangue, anchor bolts, and ironware, significantly improving practicality and reliability, effectively reducing missed detections and false detections, and providing reliable technical support for safe coal mine production.

Description

Technical Field

[0001] This application relates to the fields of computer vision and artificial intelligence technology, and in particular to a method, apparatus, equipment, medium and product for detecting foreign objects in mine conveyor belts. Background Technology

[0002] Belt conveyors are core equipment for coal transportation in underground mines. During operation, large pieces of gangue, anchor bolts, ironware, and other foreign objects can easily mix with the coal flow and enter the conveyor belt, causing damage, blockages, or even system failure, seriously affecting production safety and efficiency. Therefore, real-time and accurate automatic detection of foreign objects on the conveyor belt is crucial. However, the mining environment presents significant challenges to vision-based detection technologies, mainly including: (1) Poor image quality: Uneven lighting and dust and water vapor in the well result in low illumination, blurred details and overall dimness in the acquired images.

[0003] (2) Insufficient image resolution: Due to limitations in the performance of downhole acquisition equipment and environmental interference, the image resolution is low, and it is difficult to capture small or edge foreign object features.

[0004] (3) Complex detection targets: Foreign objects on the conveyor belt vary in size (from small pieces of gangue to long anchor rods) and shape, and are often mixed with the background of coal flow. Traditional or general target detection models are difficult to adapt to, and are prone to missed detection and false detection.

[0005] (4) Scarcity of labeled data: The underground environment is dangerous and data collection is difficult, making it difficult to obtain large-scale, high-quality labeled datasets for training robust deep learning models. Summary of the Invention

[0006] The purpose of this application is to provide a method, device, equipment, medium, and product for detecting foreign objects in mine conveyor belts. This addresses the core bottleneck problem of low input image quality caused by uneven lighting, dust interference, and equipment limitations in mines. It provides high-quality image input with uniform brightness, clear details, and sufficient resolution for subsequent high-precision detection models, laying a reliable foundation for visual perception. It also solves a series of problems inherent in general detection models in complex backgrounds, multi-scale, and irregularly shaped conveyor belt coal flow scenarios, such as insufficient feature extraction, inadequate focus on key targets, poor adaptability to special shapes, and slow and inaccurate positioning convergence. This achieves high-precision, high-recall real-time detection and accurate positioning of various foreign objects, including gangue, anchor bolts, and ironware, significantly improving practicality and reliability, effectively reducing missed and false detections, and providing reliable technical support for safe coal mine production.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for detecting foreign objects on a mine conveyor belt, including: Acquire raw images of the mine conveyor belt area; The original image is subjected to low-light image enhancement processing and super-resolution reconstruction processing in sequence to obtain a high-quality high-resolution image; Foreign object identification and localization are performed on the high-resolution image using an improved target detection model. The improved target detection model is based on a single-stage target detection network and is obtained by introducing a feature enhancement module with switchable dilated convolution to improve multi-scale feature extraction capability, introducing a focused linear attention mechanism to enhance attention to foreign object regions, introducing a dynamic serpentine convolution module to optimize feature adaptability to strip-shaped foreign objects, and using the SIOU loss function for bounding box regression optimization. The output includes the detection results, including the foreign object category label, confidence score, and location coordinates.

[0008] Optionally, the specific process of the low-light image enhancement processing includes: The original image is then input into an enhanced model based on a recurrent consistency generative adversarial network; During the encoding phase, feature extraction is enhanced through a parameterless attention mechanism; During the feature transformation stage, dilated convolution is used to expand the receptive field to integrate the illumination context; During the decoding stage, image details are recovered through a dual-channel attention mechanism.

[0009] Optionally, the specific process of the super-resolution reconstruction process includes: The low-light image enhancement process is input into a reconstruction model based on a meta-learning super-resolution network. The feature learning module of the reconstruction model adopts an adaptive dense residual structure to capture global features and embeds a multi-scale attention aggregation module to enhance the multi-scale representation of local details. The reconstruction model achieves multi-scale reconstruction through a meta-upsampling module.

[0010] Optionally, the improved target detection model is a model based on the YOLOv7 network architecture; the switchable dilated convolution feature enhancement module is set in the backbone network of the model; the focused linear attention mechanism is inserted at the end of the backbone network; and the dynamic serpentine convolution module is set in the neck network of the model.

[0011] Optionally, the foreign object detection method for mine conveyor belts further includes pre-training the improved target detection model, wherein the specific pre-training process includes: The improved object detection model is pre-trained using a large-scale general object detection dataset to obtain initial weights; Using a labeled dataset of foreign objects on mine conveyor belts, the improved target detection model loaded with the initial weights is fine-tuned and trained.

[0012] Optionally, the specific process of pre-training the improved object detection model using a large-scale general object detection dataset to obtain initial weights includes: The number of categories output by the detection head of the improved object detection model is adjusted to the total number of categories in the selected general object detection dataset; Using the training set of the general object detection dataset, the configured model is trained in multiple rounds using a composite loss function consisting of bounding box regression loss, object confidence loss, and classification loss, and the model parameters are updated by an optimizer. After training is completed, save all model parameters corresponding to the optimal performance of the model on the validation set as the initial weights. The specific process of fine-tuning the improved target detection model loaded with the initial weights using the labeled foreign object dataset from mine conveyor belts includes: The initial weights are loaded into the improved target detection model, and the number of categories output by its detection head is adjusted to the total number of foreign object categories in the foreign object dataset of the mine conveyor belt; at the same time, the training images and their annotation information of the foreign object dataset of the mine conveyor belt are loaded. The optimizer is initialized with a learning rate lower than that of the pre-training stage. The training set of the foreign object dataset of the mine conveyor belt is used to conduct supervised training on the model with initial weights using a composite loss function consisting of bounding box regression loss, target confidence loss and classification loss, and to update all or part of the model parameters. After fine-tuning the training, save the model parameters corresponding to the optimal performance on the validation set to obtain the final model suitable for foreign object detection on mine conveyor belts.

[0013] Secondly, this application provides a foreign object detection device for mine conveyor belts, comprising: The data acquisition module is used to acquire raw images of the mine conveyor belt area; An image preprocessing module, connected to the data acquisition module, is used to sequentially perform low-light enhancement and super-resolution reconstruction processing on the original image; the image preprocessing module includes a cascaded low-light image enhancement unit and a super-resolution reconstruction unit; A foreign object detection module, connected to the image preprocessing module, is used to detect foreign objects in the processed image using an improved target detection model. The output module, connected to the foreign object detection module, is used to output the detection results, which include foreign object category labels, confidence scores, and location coordinates.

[0014] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the foreign object detection method for mine conveyor belts as described above.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the foreign object detection method for mine conveyor belts described above.

[0016] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the foreign object detection method for mine conveyor belts described above.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, apparatus, equipment, medium, and product for detecting foreign objects on mine conveyor belts. By sequentially performing low-light image enhancement and super-resolution reconstruction processing on the original image, it solves the core bottleneck problem of low input image quality (low light, low resolution, and blurred details) caused by uneven lighting, dust interference, and equipment limitations in mines. This provides high-quality image input with uniform brightness, clear details, and sufficient resolution for subsequent high-precision detection models, laying a reliable foundation for visual perception. Furthermore, the improved target detection model performs foreign object identification and localization on high-resolution images, solving the problem of... In complex, multi-scale, and irregularly shaped coal flow scenarios in conveyor belts, general detection models suffer from a series of problems, including insufficient feature extraction, inadequate attention to key targets, poor adaptability to special shapes, and slow and inaccurate localization convergence. Specifically, by introducing a feature enhancement module with switchable dilated convolution, a focused linear attention mechanism, a dynamic serpentine convolution module, and the SIOU loss function, high-precision, high-recall real-time detection and accurate localization of various foreign objects such as gangue, anchor bolts, and ironware are achieved. This significantly improves practicality and reliability, effectively reduces missed detections and false detections, and provides reliable technical support for safe coal mine production. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an application environment diagram of a foreign object detection method for a mine conveyor belt according to an embodiment of this application; Figure 2 A schematic flowchart of a foreign object detection method for a mine conveyor belt provided in an embodiment of this application; Figure 3 This is a schematic diagram of the network structure of the enhancement model in one embodiment of this application; Figure 4 This is a schematic diagram of the network structure of the reconstruction model in one embodiment of this application; Figure 5 This is a schematic diagram of the network structure of an improved target detection model in one embodiment of this application; Figure 6 A schematic diagram of the functional modules of a foreign object detection device for a mine conveyor belt provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The foreign object detection method for mine conveyor belts provided in this application can be applied to, for example... Figure 1In the application environment shown, the image acquisition device 102 communicates with the computing processing device 104 via a network. The data storage system can store the video streams, image data, and model parameters that the computing processing device 104 needs to process. The data storage system can be integrated into the computing processing device 104 or set up separately on a local server in the mine or a cloud server. The image acquisition device 102 can send the acquired raw video or images of the mine conveyor belt to the computing processing device 104 in real time or on demand. After receiving the raw images, the computing processing device 104 performs serialization preprocessing on them using deep learning-based low-light image enhancement and super-resolution reconstruction. Then, the resulting high-quality images are input into an improved and trained target detection model for foreign object identification and location, ultimately generating detection results including the foreign object category, location, and confidence level. The computing processing device 104 can feed back the detection results to the monitoring terminal 106 for visual alarm purposes, or directly send commands to the control actuator 108 to trigger a cleanup operation. In addition, in some embodiments, the foreign object detection method for mine conveyor belts can also be implemented independently by edge computing devices deployed underground, completing the entire process from image acquisition to result output.

[0023] The image acquisition device 102 can be, but is not limited to, various intrinsically safe high-definition cameras for mining, explosion-proof surveillance cameras, infrared thermal imagers, etc. The computing and processing device 104 can be implemented using servers, server clusters, or cloud servers deployed on the ground, or using explosion-proof computers, edge computing gateways, AI analysis boxes, etc. deployed underground in the mine. The monitoring terminal 106 can be, but is not limited to, a large monitoring screen in a ground dispatch center, an industrial control computer, an explosion-proof display screen underground, a mobile inspection terminal, etc. The control and execution mechanism 108 can be, but is not limited to, an audible and visual alarm, an emergency stop device, a robotic arm, an automatic sorting robot, etc.

[0024] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting foreign objects on a mine conveyor belt is provided. This method is executed by computer equipment, specifically by a single computer device such as an edge computing device, a server, or a cloud server, or by a collaborative execution of an image acquisition device, an edge computing device, and a cloud server. In this embodiment, the method is applied to... Figure 1 The following description uses the computing processing device 104 as an example, including steps 201 to 204. Wherein: Step 201: Obtain the original image of the mine conveyor belt area; Step 202: Perform low-light image enhancement processing and super-resolution reconstruction processing on the original image in sequence to obtain a high-quality high-resolution image; Step 203: Foreign object identification and localization are performed on high-resolution images using an improved target detection model. The improved target detection model is based on a single-stage target detection network. It is obtained by introducing a feature enhancement module with switchable dilated convolution to improve multi-scale feature extraction capability, introducing a focused linear attention mechanism to enhance attention to foreign object regions, introducing a dynamic serpentine convolution module to optimize feature adaptation to strip-shaped foreign objects, and using the SIOU loss function for bounding box regression optimization. Step 204: Output the detection results, which include the foreign object category label, confidence score, and location coordinates.

[0025] By implementing steps 201 to 204 above, and sequentially performing low-light image enhancement and super-resolution reconstruction on the original image, the core bottleneck problem of low input image quality (low light, low resolution, and blurred details) caused by uneven lighting, dust interference, and equipment limitations in mines is solved. This provides high-quality image input with uniform brightness, clear details, and sufficient resolution for subsequent high-precision detection models, laying a reliable foundation for visual perception. Furthermore, the improved target detection model enables foreign object identification and localization in high-resolution images, solving the problem of foreign object recognition and localization in complex backgrounds and multi-scale environments. In scenarios involving coal flow in conveyor belts with varying degrees of density and irregular shapes, general detection models suffer from a series of problems, including insufficient feature extraction, inadequate attention to key targets, poor adaptability to special shapes, and slow and inaccurate localization convergence. Specifically, by introducing a feature enhancement module with switchable dilated convolution, a focused linear attention mechanism, a dynamic serpentine convolution module, and the SIOU loss function, high-precision, high-recall real-time detection and accurate localization of various foreign objects such as gangue, anchor bolts, and ironware are achieved. This significantly improves practicality and reliability, effectively reduces missed detections and false detections, and provides reliable technical support for safe coal mine production.

[0026] As an optional implementation method, the specific process of low-light image enhancement processing includes: The original image is input into an enhanced model based on a recurrent consistency generative adversarial network; During the encoding phase, feature extraction is enhanced through a parameterless attention mechanism; During the feature transformation stage, dilated convolution is used to expand the receptive field to integrate the illumination context; During the decoding stage, image details are recovered through a dual-channel attention mechanism.

[0027] In practice, the original low-light image X of the mine conveyor belt area is input into a pre-trained augmentation model (i.e., the generator G in CycleGAN). This augmentation model is a deep neural network specifically designed for image-to-image transformation. Its structure includes three core stages: encoder, transformer, and decoder, which work together to complete the mapping from the low-light domain to the normal-light domain. Its network structure is as follows: Figure 3As shown. In the encoding stage, the image is first preliminarily feature-mapped using a reflection-pad2d layer and a standard convolutional layer; subsequently, the feature maps are fed into a parameterless attention module. The core of this parameterless attention module (SimAM module) lies in defining an energy function for each neuron in the feature map, which can be expressed as: ; In the formula, Represents the energy function, when And all others When the minimum value is obtained, and They are two different values. Indicates an index in the spatial dimension. Indicates weight, This indicates the bias transformation.

[0028] Neurons with lower energy are considered more important in visual processing. This parameterless attention module calculates the energy of all neurons and assigns three-dimensional attention weights accordingly. Without introducing additional learnable parameters, it adaptively suppresses unimportant features such as background noise while enhancing the expression of key textures and edges related to foreign objects. This process effectively solves the problem of foreign object features being mixed with the background and difficult to extract in low-light images. In the feature transformation stage, each residual block is reconstructed into a brightness enhancement module. This module replaces standard convolutions with dilated convolutions with different dilation rates (e.g., 1 and 3). Dilated convolutions allow the kernel to skip some pixels when sampling the input, significantly expanding the network's receptive field without increasing the number of parameters. This enables the network to simultaneously perceive geographically distant regions in the feature map, effectively integrating the global illumination context information of the entire image. Through this fusion of multi-scale receptive fields, the network can more accurately estimate and correct global and local illumination conditions, thus robustly completing the nonlinear mapping from the low-light feature space to the normal-light feature space, laying the foundation for outputting images with uniform brightness. In the decoding stage, the feature map is upsampled through deconvolutional layers to gradually restore the spatial dimensions of the image. A dual-channel attention module (such as CBAM) is introduced along the upsampling path. This module performs the following steps sequentially: compressing the feature map along the channel dimension (using global max pooling and average pooling) and generating channel weights through a lightweight network; compressing the weighted feature map along the spatial dimension and generating a spatial weight map through convolutional operations. Through this dual weighting of channel and spatial attention, the network can adaptively recover and enhance shallow details such as foreign object edges and textures that may have been lost during encoding and downsampling, while suppressing artifacts. Finally, after a convolutional layer and a Tanh activation function, the final normal illumination enhanced image G(X) is output.

[0029] As an optional implementation method, the specific process of super-resolution reconstruction includes: The low-light image enhancement process is input into a reconstruction model based on a meta-learning super-resolution network. The feature learning module of the reconstruction model adopts an adaptive dense residual structure to capture global features and embeds a multi-scale attention aggregation module to enhance the multi-scale representation of local details. The reconstruction model achieves multi-scale reconstruction through a meta-upsampling module.

[0030] In practice, the low-light image enhancement image, along with a specified magnification factor, is input into a pre-trained reconstruction model (i.e., the Meta-SR network), whose network structure is as follows: Figure 4 As shown. The core of this reconstruction model lies in its ability to dynamically generate upsampling filters based on arbitrary scaling factors, enabling a single model to support multi-scale super-resolution. Low-resolution images first enter the feature learning module, where a shallow feature layer is quickly extracted. These shallow features are then fed into a series of adaptive dense residual modules. Each adaptive dense residual module introduces an adaptive feature enhancement module on top of the standard dense connection structure. Features are first normalized and embedded via convolution, then processed through a spatial context module (using large kernel group convolution to expand the receptive field and capture global structural information) and a feature refinement module (focusing on high and low frequency local information). The outputs of both modules are fused and processed by a multilayer perceptron to achieve adaptive global feature enhancement. Through the stacking of multiple adaptive dense residual modules, the Meta-SR network can fully model the global dependency between foreign objects and the background. At key locations in the feature learning module (such as after shallow features and before deep features), a multi-scale attention aggregation module is embedded. This module executes in parallel: it extracts multi-scale spatial features using convolutional kernels of different scales and generates a spatial weight map through pooling and attention mechanisms, emphasizing key spatial locations such as object edges and textures; it also evaluates and weights the importance of different feature channels through a channel attention mechanism, highlighting feature channels related to the object. The fusion of these two outputs significantly enhances the network's ability to focus on and express local details of objects at multiple scales (especially small scales), effectively avoiding the loss of details in deep networks. After deep feature learning, a meta-upsampling module is used for upsampling, which can be expressed by the following formula: ; In the formula, Indicates SR image Pixel value at that location, Pixels in a low-resolution (LR) image representing foreign objects. Features The pixel under the current scaling factor The filter weights, It is a feature mapping function used to calculate the value of a pixel.

[0031] For any target pixel in a high-resolution image, the meta-upsampling module first projects its coordinates back to the low-resolution feature map. Then, using a weight prediction network (a lightweight fully connected network) as input, it dynamically predicts a set of upsampling filter weights specific to that pixel. These upsampling filters are not fixed but are generated "on the fly" based on the magnification ratio and pixel location. Using these dynamically predicted filter weights, the features at the corresponding location and its neighborhood in the low-resolution feature map are weighted and summed (convolution operation) to directly calculate the pixel value of that target pixel in the high-resolution image. By traversing all pixels in the high-resolution image and repeating the above process, a complete high-resolution output image at any scale is finally synthesized. This process generates optimized upsampling kernels for different magnification ratios, thus better preserving and reconstructing the edge and texture details of foreign objects while magnifying the image.

[0032] As an optional implementation, the improved object detection model is a model based on the YOLOv7 network architecture; the feature enhancement module with switchable dilated convolution is set in the backbone network of the model; the focused linear attention mechanism is inserted at the end of the backbone network; and the dynamic serpentine convolution module is set in the neck network of the model.

[0033] In practical implementation, the improved target detection model uses the standard YOLOv7 network as the baseline architecture. This architecture includes an input network, a backbone network, a neck network, and a head, and its network structure is as follows: Figure 5 As shown. The improved object detection model receives a high-resolution conveyor belt image preprocessed (low-light enhancement and super-resolution reconstruction) as input. A feature enhancement module is integrated into the backbone network, embedded in the mid-to-deep layers. Its main function is to adaptively expand the network's receptive field and enhance feature extraction capabilities to better capture the features of multi-scale foreign objects, especially improving sensitivity to small-sized foreign objects. Before the deep feature map is output to the neck network after feature extraction is completed in the backbone network, a focusing linear attention mechanism is inserted, which can be expressed by the formula: ; In the formula, Indicates element-wise calculation power of p, using By adjusting the feature orientation, it can be proven that... This indicates that the norm of the features remains unchanged after mapping, while the feature orientation is adjusted, which can achieve the differential distribution between similar and dissimilar query-key pairs.

[0034] This focused linear attention mechanism is applied to the deep, low-resolution feature maps output by the backbone network. Its main function is to achieve computationally efficient global attention modeling, enabling the network to focus on potential foreign object regions in the image and effectively suppress interference from complex backgrounds such as conveyor belt textures and coal flows. Specifically, it performs a non-linear transformation on the Query and Key features through a mapping function (such as element-wise exponentiation) to increase the difference between relevant and irrelevant feature pairs, thereby achieving "focusing". At the same time, it compensates for the lack of feature diversity that may be caused by linear attention by adding depthwise separable convolution. In the feature fusion path of the neck network (such as the PANet path aggregation network structure), some standard convolution operations are replaced with dynamic serpentine convolution modules. These dynamic serpentine convolution modules are mainly used in the stage of fusing and re-extracting feature maps from different levels of the backbone network. Their core function is to enhance the network's ability to extract and adapt to the geometric features of irregularly shaped objects (such as anchor rods and wires), specifically by adaptively performing convolution operations along the extension direction of the irregular object, thereby more accurately extracting its continuous structural features and generating feature representations that better fit the actual shape of the target. This effectively alleviates the shortcomings of standard rectangular convolution kernels in describing slender objects. During the model training phase, the SIOU loss function is used instead of the default CIoU loss function in YOLOv7 for bounding box regression, which can be expressed as: ; In the formula, Indicates angle loss, Indicates the loss by intersection and union. Indicates distance loss. Indicates shape loss, , , , The degree of attention paid to shape loss is controlled, where w and h are the width and height of the true bounding box.

[0035] SIOU, in addition to considering overlap area, center-to-center distance, and aspect ratio, introduces an additional angular cost between bounding boxes. It accelerates convergence by minimizing the vector angle between the predicted and ground truth boxes. During backpropagation in training, the SIOU loss calculates gradients based on the angle, distance, shape, and IoU between the predicted and ground truth boxes, guiding the network to adjust the position and size of the predicted boxes faster and more accurately.

[0036] In another exemplary embodiment of this application, in order to ensure high-precision, high-recall real-time detection and accurate positioning of various foreign objects such as gangue, anchor bolts, and ironware, the method may further include pre-training an improved target detection model before step 203. The specific pre-training process includes: The improved object detection model is pre-trained using a large-scale general object detection dataset to obtain initial weights; Using a labeled dataset of foreign objects on mine conveyor belts, we fine-tuned the training of an improved object detection model with initial weights.

[0037] As an optional implementation, the process of pre-training the improved object detection model using a large-scale general object detection dataset to obtain initial weights includes: Adjust the number of categories output by the detection head of the improved object detection model to the total number of categories in the selected general object detection dataset; Using the training set of a general object detection dataset, the configured model is trained in multiple rounds of iteration using a composite loss function consisting of bounding box regression loss, object confidence loss, and classification loss, and the model parameters are updated through an optimizer. After training is complete, save all model parameters corresponding to the optimal performance of the model on the validation set as initial weights.

[0038] In practice, a publicly available large-scale general object detection dataset is selected as the pre-training data source, such as MSCOCO or PASCAL VOC. Taking the COCO dataset as an example, it contains 80 everyday object categories and over 200,000 labeled images, covering a wide range of scenes and perspectives. The dataset's annotation format (e.g., JSON) is converted to the format required for training the object detection model (e.g., YOLO format txt files). The number of channels in the convolutional layers used for classification output in the detection head of the improved object detection model is modified. Specifically, the number of output channels is adjusted from the number of categories targeting foreign objects in the mine (e.g., 3 categories: gangue, anchor bolts, ironware) to the total number of categories in the selected general dataset (e.g., 80 categories in the COCO dataset). The structure and parameters of the model's backbone network, neck network, and all other improved modules (FEM, FT, EDSC) remain unchanged in their initial random state or standard initialization. Data augmentation strategies were implemented during training to improve model robustness, including random horizontal flipping, random scaling (e.g., 0.5x to 1.5x), mosaic enhancement, and color space perturbations (hue, saturation, and brightness adjustments). Simultaneously, all input images were normalized and scaled to a uniform network input size (e.g., 640x640 pixels). The object detection model was optimized end-to-end using its inherent composite loss function. This loss function consists of three parts: bounding box regression loss, object confidence loss, and classification loss. The bounding box regression loss uses the aforementioned improved SIOU Loss to accurately optimize the position and size of the predicted boxes. The object confidence loss uses binary cross-entropy loss to determine whether each predicted box contains an object. The classification loss uses multivariate cross-entropy loss to correctly classify the detected objects. An adaptive moment estimation optimizer, such as Adam or AdamW, was selected, and weight decay was implemented to prevent overfitting. The learning rate strategy employs a cosine annealing or a cosine annealing learning rate scheduler with hot restart, setting an initial learning rate that gradually decreases with each training epoch. The total number of training epochs (e.g., 300 epochs) is set based on the dataset size and computational resources, and an appropriate batch size is determined to fully utilize GPU memory. The general dataset is divided into training and validation sets. Batch images are input into the object detection model, passing through the backbone network, neck network, and detection head, outputting predicted bounding boxes, confidence scores, and class probabilities. The composite loss is calculated based on the prediction results and ground truth annotations. The gradient of the loss with respect to all trainable parameters of the model is calculated using backpropagation. The optimizer uses the calculated gradients to update the model parameters, minimizing the total loss. After every few training epochs, model performance is evaluated on an independent validation set, with the mean accuracy (MA) as the key metric. Changes in this metric are monitored to determine model convergence and select optimal weights.After all training epochs are completed, or after training is terminated according to an early stopping strategy (e.g., the validation set mAP does not improve for several consecutive epochs), select the model state corresponding to the epoch with the highest mAP on the validation set. At this point, the parameters of the object detection model have been fully optimized on large-scale general data, possessing excellent general object detection capabilities. Save all parameters (i.e., weights and biases) of this optimal model state into a separate file, which contains the initial weights for each layer of the model.

[0039] The specific process of fine-tuning the improved target detection model with initial weights using a labeled dataset of foreign objects on mine conveyor belts includes: The initial weights are loaded into the improved object detection model, and the number of categories output by its detection head is adjusted to the total number of foreign object categories in the mine conveyor belt foreign object dataset; at the same time, the training images and their annotation information of the mine conveyor belt foreign object dataset are loaded. The optimizer is initialized with a learning rate lower than that of the pre-training stage. The training set of the foreign object dataset of the mine conveyor belt is used. The model with initial weights is trained in a supervised manner using a composite loss function consisting of bounding box regression loss, target confidence loss and classification loss, and all or part of the model parameters are updated. After fine-tuning the training, save the model parameters corresponding to the optimal performance on the validation set to obtain the final model suitable for foreign object detection on mine conveyor belts.

[0040] In practice, a labeled dataset of foreign objects on a mine conveyor belt is loaded. This dataset contains a large number of images collected from the mine site. The foreign objects (such as gangue, anchor bolts, and ironware) in each image are labeled with bounding boxes and category labels. The dataset is divided into training set, validation set, and test set in a ratio (e.g., 8:1:1).

[0041] The initial weight file obtained in the pre-training phase is loaded into the improved object detection model. The number of output channels of the convolutional layer used for classification output in the detection head is readjusted from the total number of categories in the general dataset used during pre-training (e.g., 80 categories) to the total number of foreign object categories (e.g., 3 categories: gangue, anchor bolts, and ironware). Considering the characteristics of mine data, a data augmentation strategy is configured for the fine-tuning phase. In addition to conventional random flipping and scaling, augmentations specific to the mine scene are added, such as simulating dust noise and local brightness changes, to improve the model's robustness to changes in the underground environment. The input image is also normalized and scaled to a uniform size (e.g., 640x640). The fine-tuning phase continues to use the same composite loss function as the pre-training phase, i.e., a composite loss function consisting of a weighted sum of SIOU bounding box regression loss, object confidence loss, and classification loss. The learning rate is lower than the initial learning rate in the pre-training phase. A lower learning rate ensures that when updating parameters, the valuable general features learned in the pre-training weights are not destroyed too quickly or too much, allowing for fine-tuning. The optimizer uses either the Adam or AdamW optimizer, but its momentum state is reinitialized. A gentler learning rate descent strategy is employed, such as decaying at fixed epochs or automatically decaying when validation set performance plateaus. Parameter updates use differential learning rates or partial parameter freezing strategies; for example, training only the head and neck detection networks and parts of the deep backbone while freezing the parameters of the shallow backbone (because it contains more general low-level features), or training all parameters but setting different learning rates for different layers. Training images from the mine dataset are batched into the improved object detection model, the loss is calculated, and backpropagation is performed. Since the model parameters are initialized by pre-trained weights, gradient descent focuses on how to better adapt these general features to the specific task of mine foreign objects. Model performance is evaluated on the validation set of the mine dataset after every few training epochs. Monitoring metrics include mean precision for mine foreign objects, mean precision for each class, and recall. To prevent overfitting, an early stopping strategy is implemented, automatically stopping training when the validation set mAP no longer improves over several consecutive epochs. After fine-tuning the training, the model corresponding to the training round with the best overall performance (usually the highest mAP) on the mine validation set is selected. At this point, the improved object detection model retains general visual knowledge while its parameters have been fully optimized for the mine foreign object detection task. All parameters of this optimal model state are saved as the final model file, which is the final model that can be directly deployed for real-time foreign object detection in the mine.

[0042] This application also provides an application scenario in which the aforementioned foreign object detection method for mine conveyor belts is applied. Specifically, the foreign object detection method for mine conveyor belts provided in this embodiment can be applied in the intelligent safety production and transportation assurance scenario of coal mines. Coal enters the transportation chain from the mining stage and is transferred to the surface through a relay of multiple belt conveyors. During this process, foreign objects mixed into the conveyor belt (such as gangue, anchor bolts, and ironware) are the main risk sources causing conveyor belt damage, system shutdown, and even safety accidents. The foreign object detection method for mine conveyor belts provided in this embodiment belongs to the intelligent safety monitoring link of the conveyor belt transportation process. Specifically, in the real-time monitoring of conveyor belt operation, based on the fully automatic intelligent detection and manual inspection verification of the method, real-time identification, location, and early warning of foreign objects can be achieved. This provides a reliable basis for the safety production management system to perceive and handle abnormal events, thereby forming a closed-loop safety management link of "perception-analysis-early warning-handling", comprehensively improving the intelligence level and inherent safety level of the mine transportation system.

[0043] Based on the same inventive concept, this application also provides a foreign object detection device for mine conveyor belts to implement the aforementioned method for detecting foreign objects in mine conveyor belts. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the foreign object detection device for mine conveyor belts provided below can be found in the limitations of the method for detecting foreign objects in mine conveyor belts described above, and will not be repeated here.

[0044] In one exemplary embodiment, such as Figure 6 As shown, a foreign object detection device for a mine conveyor belt is provided, comprising: The data acquisition module is used to acquire raw images of the mine conveyor belt area; The image preprocessing module, connected to the data acquisition module, is used to sequentially perform low-light enhancement and super-resolution reconstruction on the original image; the image preprocessing module includes a cascaded low-light image enhancement unit and a super-resolution reconstruction unit. The foreign object detection module, connected to the image preprocessing module, is used to detect foreign objects in the processed image using an improved target detection model. The output module, connected to the foreign object detection module, is used to output detection results including foreign object category labels, confidence scores, and location coordinates.

[0045] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores video streams, image data, and model parameters that need to be processed. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When executed by the processor, the computer program implements a method for detecting foreign objects on a mine conveyor belt.

[0046] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0047] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0048] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0049] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0050] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0051] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0052] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting foreign objects on a mine conveyor belt, characterized in that, The method for detecting foreign objects on mine conveyor belts includes: Acquire raw images of the mine conveyor belt area; The original image is subjected to low-light image enhancement processing and super-resolution reconstruction processing in sequence to obtain a high-quality high-resolution image; Foreign object identification and localization are performed on the high-resolution image using an improved target detection model. The improved target detection model is based on a single-stage target detection network and is obtained by introducing a feature enhancement module with switchable dilated convolution to improve multi-scale feature extraction capability, introducing a focused linear attention mechanism to enhance attention to foreign object regions, introducing a dynamic serpentine convolution module to optimize feature adaptability to strip-shaped foreign objects, and using the SIOU loss function for bounding box regression optimization. The output includes the detection results, including the foreign object category label, confidence score, and location coordinates.

2. The method for detecting foreign objects in a mine conveyor belt according to claim 1, characterized in that, The specific process of low-light image enhancement includes: The original image is input into an enhanced model based on a recurrent consistency generative adversarial network; During the encoding phase, feature extraction is enhanced through a parameterless attention mechanism; During the feature transformation stage, dilated convolution is used to expand the receptive field to integrate the illumination context; During the decoding stage, image details are recovered through a dual-channel attention mechanism.

3. The method for detecting foreign objects in a mine conveyor belt according to claim 1, characterized in that, The specific process of the super-resolution reconstruction process includes: The low-light image enhancement process is input into a reconstruction model based on a meta-learning super-resolution network. The feature learning module of the reconstruction model adopts an adaptive dense residual structure to capture global features and embeds a multi-scale attention aggregation module to enhance the multi-scale representation of local details. The reconstruction model achieves multi-scale reconstruction through a meta-upsampling module.

4. The method for detecting foreign objects in a mine conveyor belt according to claim 1, characterized in that, The improved target detection model is based on the YOLOv7 network architecture; the switchable dilated convolution feature enhancement module is set in the backbone network of the model; The focused linear attention mechanism is inserted at the end of the backbone network; the dynamic serpentine convolution module is set in the neck network of the model.

5. The method for detecting foreign objects in a mine conveyor belt according to claim 1, characterized in that, It also includes pre-training the improved object detection model, the specific process of which includes: The improved object detection model is pre-trained using a large-scale general object detection dataset to obtain initial weights; Using a labeled dataset of foreign objects on mine conveyor belts, the improved target detection model loaded with the initial weights is fine-tuned and trained.

6. The method for detecting foreign objects in a mine conveyor belt according to claim 5, characterized in that, The specific process of pre-training the improved object detection model using a large-scale general object detection dataset to obtain initial weights includes: The number of categories output by the detection head of the improved object detection model is adjusted to the total number of categories in the selected general object detection dataset; Using the training set of the general object detection dataset, the configured model is trained in multiple rounds using a composite loss function consisting of bounding box regression loss, object confidence loss, and classification loss, and the model parameters are updated by an optimizer. After training is completed, save all model parameters corresponding to the optimal performance of the model on the validation set as the initial weights. The specific process of fine-tuning the improved target detection model loaded with the initial weights using the labeled foreign object dataset from mine conveyor belts includes: The initial weights are loaded into the improved target detection model, and the number of categories output by its detection head is adjusted to the total number of foreign object categories in the foreign object dataset of the mine conveyor belt; at the same time, the training images and their annotation information of the foreign object dataset of the mine conveyor belt are loaded. The optimizer is initialized with a learning rate lower than that of the pre-training stage. The training set of the foreign object dataset of the mine conveyor belt is used to conduct supervised training on the model with initial weights using a composite loss function consisting of bounding box regression loss, target confidence loss and classification loss, and to update all or part of the model parameters. After fine-tuning the training, save the model parameters corresponding to the optimal performance on the validation set to obtain the final model suitable for foreign object detection on mine conveyor belts.

7. A foreign object detection device for mine conveyor belts, characterized in that, For implementing the foreign object detection method for mine conveyor belts as described in any one of claims 1-6, the foreign object detection device for mine conveyor belts comprises: The data acquisition module is used to acquire raw images of the mine conveyor belt area; An image preprocessing module, connected to the data acquisition module, is used to sequentially perform low-light enhancement and super-resolution reconstruction processing on the original image; the image preprocessing module includes a cascaded low-light image enhancement unit and a super-resolution reconstruction unit; A foreign object detection module, connected to the image preprocessing module, is used to detect foreign objects in the processed image using an improved target detection model. The result output module is connected to the foreign object detection module and is used to output the detection results, which include foreign object category labels, confidence scores, and location coordinates.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the foreign object detection method for mine conveyor belts according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the foreign object detection method for mine conveyor belts as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the foreign object detection method for mine conveyor belts as described in any one of claims 1-6.

Images