A coal mine belt foreign matter detection method and system based on N-NanoDet

Abstract

The application discloses a coal mine belt foreign matter detection method and system based on N-NanoDet, and belongs to the technical field of coal mine safety monitoring. The method comprises the following steps: an image enhancement step, based on a physical prior memory storage and memory enhancement guidance mechanism, historical frame prior information is used to restore degraded images, and an edge constraint module is used to perform fidelity processing on the edges of objects; a feature extraction and fusion step, a multi-level foreign matter feature is extracted by using a ShuffleNet_V2, a bidirectional multi-scale feature pyramid is constructed by combining an FPN and a PAN to perform feature fusion; and a target detection step, an N-NanoDet detection network based on an anchor-free mechanism is used to regress and locate foreign matters by improving an SIoU loss function. The application solves the problems of image degradation, difficulty in detecting small foreign matters and limited edge computing power in a low-illumination and high-dust environment underground, and has the characteristics of high detection precision, fast reasoning speed and efficient edge deployment.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent mine safety monitoring technology, specifically relating to a method and system for detecting foreign objects on coal mine conveyor belts based on N-NanoDet. Background Technology

[0002] Foreign objects mixed on the surface of underground conveyor belts in coal mines are a major cause of serious equipment accidents such as longitudinal tearing, belt misalignment, and even sudden belt breakage. Because such accidents can paralyze the entire conveyor line within seconds, accompanied by high maintenance costs and safety risks, accurate and rapid identification and early warning of these foreign objects during belt operation has become a core element in ensuring the inherent safety of mine conveyor systems. However, the extreme complexity of actual underground working conditions and the technological gap between existing detection methods have long presented practical challenges to achieving this safety goal.

[0003] Specifically, existing industrial monitoring and visual inspection solutions face three major challenges when deployed underground. First, there's the conflict between environmental degradation and image quality: low illumination, coal dust dispersion, and severe water mist interference in underground environments result in monitoring images exhibiting nonlinear illumination decay, global contrast compression, and noise-induced target edge submersion. This significantly weakens the effectiveness of traditional morphological segmentation and shallow feature extraction methods, leading to alternating false positives and false negatives. Second, there's the conflict between foreign object features and detection accuracy: foreign objects in the context of high-speed conveyor belt movement vary greatly in size, from centimeter-sized bolts to half-meter-square pieces of gangue. They often appear as sparse, tiny targets in the overall image, with a very low pixel count, making it difficult for deep convolutional networks to respond stably and exacerbating feature blurring and location drift. Third, there's the conflict between model complexity and edge computing power: high-performance deep learning detectors rely on massive parameter stacking and floating-point operations, while mining edge embedded devices are constrained by explosion-proof packaging, power consumption, heat dissipation, and cost, resulting in extremely limited computing resources that cannot accommodate the continuous operation of high-precision models.

[0004] However, a deeper examination of existing technical solutions reveals more fundamental flaws in their underlying processing logic. At the image preprocessing level, existing algorithms largely rely on traditional histogram equalization or fixed Retinex enhancement models, lacking the ability to adapt to non-uniform dust and dynamic lighting changes in underground environments. This often results in weak foreign object signals being submerged in the enhanced background noise while simultaneously increasing overall brightness. Regarding feature extraction mechanisms, mainstream lightweight networks are limited by the fixed receptive field, lacking the ability to model the dynamic deformation of strip-shaped foreign objects that undergo deformation, rotation, or partial occlusion under high-speed motion. The isovariant design of convolutional kernels leads to severe feature misalignment when dealing with non-rigid foreign objects, causing the semantic confidence distribution to tend towards randomization. Furthermore, existing solutions are too crude in utilizing temporal information, relying solely on single-frame spatial features for discrimination, ignoring strong discriminative cues such as the continuity of the foreign object's trajectory and abrupt changes in velocity between adjacent frames. This reasoning paradigm, when faced with similar grayscale materials like gangue or semi-buried iron objects, suffers from frequent state flips due to the lack of spatiotemporal consistency constraints. Summary of the Invention

[0005] This invention aims to propose a method and system for detecting foreign objects in coal mine conveyor belts based on N-NanoDet. By synergistically optimizing image enhancement in low-light dust environments, multi-scale feature fusion of small foreign objects, and lightweight deployment at the edge, it overcomes the bottleneck of visual detection under complex underground working conditions and ensures real-time safety early warning for coal mine conveyor belt systems.

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

[0007] A method for detecting foreign objects on coal mine conveyor belts based on N-NanoDet, characterized by comprising:

[0008] 1) Image enhancement step: Obtain the original image frames of the coal mine conveyor belt monitoring video, and perform enhancement processing on the original image frames based on physical priors and edge constraints to obtain an enhanced image. The enhancement processing encodes and compresses the prior information of historical frames and stores it in the local prior memory matrix through the physical prior memory storage module. The memory enhancement physical prior guidance module uses the prior memory matrix to enhance the prior features of the current frame. Then, the multi-scale scene radiation recovery module captures the spatiotemporal correspondence between frames to restore the degraded image. Finally, the edge constraint module performs fidelity processing on the object edge details of the restored image.

[0009] 2) Feature extraction and fusion step: The enhanced image is input into the lightweight feature extraction network ShuffleNet_V2 to extract multi-level foreign object features, and a bidirectional multi-scale feature pyramid is constructed using FPN and PAN. The multi-level foreign object features are then fused to obtain fused features.

[0010] 3) Target detection step: The fused features are input into the N-NanoDet detection network based on the anchor box mechanism, and regression is performed through the loss function to identify and locate foreign objects in the enhanced image.

[0011] This invention introduces a priori memory storage and retrieval mechanism for historical frames, enabling the current frame to be enhanced using accumulated prior information from long sequences, thus overcoming the limitation of insufficient enhancement information utilization in a single frame. The edge constraint module performs gradient domain fidelity processing on the restored image, avoiding edge loss caused by deviations in atmospheric light scattering model parameter estimation. The three steps form a virtuous cycle of "high-quality input → efficient feature expression → accurate target regression," improving detection accuracy in low-light and high-dust environments.

[0012] The optimized image enhancement steps specifically include:

[0013] 1) Use the ConvNeXt backbone network to extract multi-scale input features, and encode and compress the transmission map and atmospheric light prior information into prior memory tokens and save them to the local prior memory matrix through the dense residual feature fusion structure in the physical prior memory storage module.

[0014] 2) The similarity weight between the current frame and the historical prior memory matrix is ​​calculated by the memory-enhanced physical prior guidance module, which guides the attention mechanism to generate prior-guided scene features. The spatiotemporal deformable attention block of the multi-scale scene radiation recovery module is used to capture the spatiotemporal correspondence between frames, so as to output a clear image with dehazing and blurring.

[0015] 3) The color image is converted to grayscale using an edge constraint module based on dynamic adjustment of edge intensity. Then, the Sobel operator is used to calculate the gradients in the horizontal and vertical directions to obtain the edge amplitude of each pixel. Finally, the image is weighted and fused based on the edge amplitude to preserve the edge details of the object.

[0016] In the above scheme, the ConvNeXt backbone network captures the long-range dependencies of dust scattering; the physical prior memory storage module encodes prior information into compressed memory tokens, reducing memory usage while maintaining spatial structure; the memory enhancement guidance module uses cross-frame similarity weight calculation to enable the current frame to adaptively acquire historical valid information; and the multi-scale scene radiation restoration module uses spatiotemporally deformable attention blocks to capture inter-frame correspondences, suppress jitter and blur, and output clear and stable enhanced images.

[0017] The optimized physical prior memory storage module employs a 3×3 convolutional kernel and a 4-layer deep backbone branch, as well as a 3×3 convolutional kernel and a 2-layer deep mask branch, for the initial prior features. The information is learned through the encoder and backbone network, including prior information such as transmission maps and atmospheric light, where H, W, and C represent the feature height, width, and number of channels, respectively. The mask branch information enhances the initial prior feature information through additive fusion and fuses multi-level features through dense connections to output the key prior feature of the current frame. R N×D×C The key prior features are processed by two pooling layers to obtain a prior compressed memory token p. R D×C Where D represents the token length, the prior compressed memory token is stored in the local memory matrix. The main branch and the mask branch achieve fine modulation of prior features through additive fusion, and dense connections alleviate gradient vanishing and enhance representation capabilities; key prior features are compressed into low-dimensional memory tokens through two pooling layers, achieving optimal storage efficiency while retaining core prior information, providing a feasible solution for long-sequence prior memory storage of edge devices.

[0018] In the optimized memory-enhanced physical prior guidance module, the similarity weight S between the key memory feature key of the current frame and the prior memory matrix K is calculated as follows:

[0019] S = softmax ( )

[0020] Here, c1 serves as a normalization factor to align the dimensions of key and K. A similarity matrix is ​​obtained by comparing the similarity between the current frame and the memory. The similarity weights then guide the attention mechanism to enhance the current prior features, specifically as follows:

[0021] P = softmax( )S

[0022] Wherein, the query vector Q is obtained by using the initial prior features Expanding the diagram, parameter c2 is used as a normalization factor to align the dimensions of Q and K, ultimately yielding the memory-enhanced prior feature P. The enhanced prior feature is concatenated with the initial scene feature as input, and multiple convolutional layers generate prior-guided scene features. This scheme adaptively evaluates the contribution of each historical token by calculating the similarity weights between the current frame and the historical memory matrix, avoiding interference from invalid information. The normalization factor ensures computational stability, and residual connections preserve the integrity of the original features. The concatenation of the enhanced prior feature with the initial scene feature generates prior-guided scene features, achieving explicit guidance of scene reconstruction by physical priors and improving the ability to identify dusty areas.

[0023] In the optimized multi-scale scene radiometric reconstruction module, adjacent frame features are divided into multiple sets with different ranges to explore cues at different time intervals. Connectivity features of different ranges are sent to a shared spatiotemporally deformable attention block. Within the spatiotemporally deformable attention block, features of a given connection are... and normalized initial spacetime flow First, calculate the initial alignment feature map. :

[0024] = S( , )

[0025] Where S represents a differentiable spatiotemporal sampling operation; then, the refined spatiotemporal flow is obtained by calculating the flow offset residual. :

[0026] = ( )+

[0027] in It is a lightweight offset network composed of convolutional layers; cross attention extracts temporal information from alignment features. Used as a query for the target frame, alignment features As keys and values, range characteristics The calculation is as follows:

[0028] = W MSA(Q,K,V)

[0029] Alignment features of different ranges are aggregated under the guidance of prior features. The concatenated range features are regarded as keys and values, multiplied by the target frame query features to calculate scene attention weights. The prior features are regarded as queries and the aligned prior features are used as attention keys to calculate prior attention weights. The final attention weight is the sum of scene attention weights and prior attention weights. After Softmax normalization, the range features are weighted and aggregated.

[0030] The above scheme divides adjacent frames into different range sets for separate processing, while capturing the spatiotemporal dependencies of short-term rapid changes and long-term slow evolution; deformable attention blocks achieve precise inter-frame alignment through differentiable sampling and flow offset residuals, compensating for camera shake and belt movement offset; prior-guided multi-range aggregation calculates weights from both scene radiosity and physical priors, making the fusion subject to both data-driven and physical model constraints, thus improving the clarity and stability of the output frames.

[0031] In the optimized edge constraint module based on dynamic adjustment of edge intensity, the input color image is converted to grayscale, and the R, G, and B channels of each pixel are added together and then divided by 3 to obtain the grayscale pixel value.

[0032] Edge detection is computed using the Sobel operator; the Sobel convolution kernel is used to detect horizontal edges. x for:

[0033] Sobel x =

[0034] Sobel convolution kernel for detecting vertical edges y for:

[0035] Sobel y =

[0036] The formula for calculating the edge amplitude EM of each pixel is:

[0037] EM =

[0038] Among them, G x and G y The horizontal and vertical gradients of the image are obtained through convolution operations, respectively; the edge constraint image is calculated using a weighted formula.

[0039] E=E output (1) EM) E raw EM

[0040] Among them, E output For grayscale images, E raw This is the original color image.

[0041] This paper addresses the issue of edge loss caused by biased estimation of atmospheric light scattering model parameters in non-uniform dust scenarios. It achieves adaptive and faithful edge detail preservation through grayscale conversion, Sobel edge detection, and weighted fusion. The edge amplitude EM provides an objective intensity measure for each pixel. The weighted fusion formula preserves the original details in areas with severe edges and applies smoothing constraints in smooth areas to avoid noise amplification. It is especially suitable for scenarios with uneven dust concentration distribution and complex edge changes in underground wells.

[0042] The optimized feature extraction and fusion steps specifically include: using the ShuffleNet_V2 network as the backbone network, the input feature map channels are divided into left and right branches through a channel partitioning mechanism. The left branch does not undergo convolution, while the right branch is downsampled through continuous convolution and depthwise separable convolution. The output feature maps of the two branches are then combined through channel merging and channel shuffling mechanisms to extract the three-layer features of the object's edge, shape, and location layer by layer. The extracted multi-layer features are input into the FPN, where a top-down upsampling operation aligns and fuses the deep low-resolution features with the shallow high-resolution features. The fusion formula is as follows:

[0043] P l = UpSample(C l )+C l+1

[0044] Among them, P l For the characteristics of the first layer of the pyramid, C l For the corresponding layer features of the backbone network, UpSample is the upsampling operation; the features processed by FPN flow to PAN, and the shallow layer's location information flows back to the deep layer through bottom-up path aggregation. The path aggregation operation formula is:

[0045] P agg = Conv2D(P l )

[0046] Among them, P agg P represents the features after path aggregation. l Feature maps are extracted at different levels; finally, a weighted fusion mechanism is used to match the weights of features at different scales to highlight foreign object features. The weighted fusion formula is as follows:

[0047] F final =

[0048] in, The weights of the features at each scale are... By generating feature maps at different scales, the final multi-scale fused feature map is obtained.

[0049] In the above scheme, ShuffleNet_V2 extracts edge, shape and position features layer by layer through channel division and shuffling; FPN transmits deep semantics from top to bottom to enhance the shallow classification and discrimination ability; PAN feeds back shallow positioning information from bottom to top to make up for the lack of deep spatial resolution; bidirectional information flow realizes deep interaction between semantics and position; learnable weighted fusion adaptively allocates scale weights to highlight foreign object response and suppress background interference, thereby improving the accuracy of multi-scale foreign object detection.

[0050] The optimized target detection step specifically includes:

[0051] The N-NanoDet model employs an anchor-free mechanism to eliminate the dependence on predefined anchor boxes, and uses a convolutional network to directly perform position regression prediction on the centrality, width, height, and category of foreign objects on the belt.

[0052] Depthwise separable convolutions are applied to the model to reduce the computational cost and number of parameters. Features at different levels are input into the Head module through feature concatenation operations, and finally the bounding boxes of each detected target are output.

[0053] The loss function is an improved loss function, with a total loss L. SIoU Represented as:

[0054] L SIoU =1 IoU+

[0055] in, For the overlapping area loss, To account for distance loss, the object's position offset was calculated and the effect of angle on position was corrected. The aspect ratio difference loss measures the difference in the aspect ratio of the target bounding box, improving the prediction accuracy of the target bounding box shape. In this scheme, the anchor-free mechanism eliminates the dependence on predefined anchor boxes, avoids the influence of hyperparameter sensitivity, and reduces computational and memory usage; depthwise separable convolution compresses the number of parameters; and the improved loss function simultaneously constrains the overlapping area, center point distance, and aspect ratio. This triple joint optimization enables the model to achieve accurate localization and bounding box regression even in low-quality images.

[0056] The optimization also includes a dataset synthesis step for building the training dataset:

[0057] The DPT Hybrid depth estimation model is used to perform pixel-by-pixel depth prediction for each frame of input image to obtain scene geometric structure information, and a spatially distributed dust concentration density map is constructed through depth inversion.

[0058] According to the atmospheric scattering model I(x) = J(x)·t(x) + A·(1 t(x)) is used to perform fogging synthesis on a sharp image, where J(x) is the sharp image, t(x) is the transmittance, and A is the atmospheric light;

[0059] The transmittance t(x) is dynamically adjusted according to the dust concentration, t(x) = exp(-β·d(x)·C(x)), where d(x) is the distance to the light source, C(x) is the dust concentration at that location, and β is the dust attenuation coefficient;

[0060] A joint optimization mechanism of bilateral filtering and guided filtering is introduced to smooth the transmission map. The atmospheric light A(x) is adaptively estimated using a dark channel prior method, and the illumination distribution is dynamically corrected by combining local brightness and dust concentration changes. For each frame, three fogged image samples with different fog density levels (low, medium, and high) are synthesized to cover multi-scale visibility variations in the actual environment. In this invention, DPT Hybrid predicts depth pixel-by-pixel and constructs a spatially distributed dust concentration density map, which is closer to the real physical environment downhole. Transmittance considers both concentration and depth simultaneously, achieving dynamic spatial adjustment. The joint optimization of bilateral filtering and guided filtering preserves edge details. The dark channel prior adaptively estimates atmospheric light and corrects the illumination distribution. The multi-level synthesis of low, medium, and high levels covers the complete range of visibility variations, enhancing the physical realism and diversity of the training data.

[0061] This invention also introduces a foreign object detection system for coal mine conveyor belts based on N-NanoDet. Using the aforementioned method, this system includes an industrial camera, an edge computing box, an alarm, a microswitch, and a monitoring platform. The industrial camera is installed above the conveyor belt chute to acquire real-time images of the belt. The edge computing box uses an ARM architecture processor and is pre-installed with an Android operating system for processing real-time images. When the edge computing box detects a foreign object on the belt surface, it controls the alarm to sound an alarm via the microswitch, and simultaneously uploads and saves the processing results and image data to the monitoring platform. In this solution, the industrial camera is installed above the chute, achieving real-time coverage of the entire belt cross-section; the combination of ARM architecture and Android system meets the explosion-proof and low-power requirements of mining equipment; the edge computing box runs the complete detection algorithm locally, avoiding the bandwidth pressure and latency of cloud transmission; the microswitch automatically controls the alarm and synchronously uploads data to the monitoring platform, forming a closed-loop control system. This provides a complete solution for underground conveyor belt transportation safety monitoring that combines real-time performance, reliability, and engineering feasibility.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. This invention addresses the problem of non-uniform scattering and sudden drop in illuminance caused by high concentrations of dust in underground mines. It abandons traditional empirical enhancement strategies and introduces a preprocessing mechanism that integrates a physical atmospheric scattering model and edge preservation constraints. This mechanism applies local gradient regularization constraints to the transmittance map, which removes the fogging and blurring caused by suspended particles while preserving weak edge information at the junction of foreign objects and conveyor belts. This avoids the atmospheric light parameter estimation bias and key contour loss caused by the failure of local block assumptions in the classic dark channel prior algorithm. This design achieves a balance between visual fidelity and structural integrity in the enhanced image, providing high-quality data input for subsequent feature extraction and thus curbing the side effects of noise amplification.

[0064] 2. This invention overcomes the technical limitations of existing lightweight networks that are slow to respond to small targets. By using ShuffleNet_V2 as the backbone feature extraction network and introducing a feature pyramid architecture that combines bidirectional FPN and PAN, it achieves multi-directional and multi-level deep interaction between deep semantic information and shallow spatial localization information. This architecture effectively bridges the representation gap between high semantic features and high-resolution features, enabling the network to maintain a stable activation response to tiny foreign objects with extremely large scales and irregular shapes even against complex coal dust backgrounds. Simultaneously, this invention, through bidirectional cross-layer connections and adaptive scale redistribution mechanisms, suppresses the interference of background textures on foreign object features, ensuring that sparse targets with low pixel ratios can still be accurately locked and identified in the global receptive field, thus improving the system's full-scene perception capability for dangerous foreign objects.

[0065] 3. This invention achieves optimal synergy between accuracy and efficiency through a systematic lightweight design. On one hand, it employs a joint strategy of anchor-free detection and depthwise separable convolution, reducing redundant computational costs and storage bandwidth pressure associated with traditional anchor point matching, thus simultaneously reducing the number of model parameters and floating-point operations. On the other hand, it introduces the SIoU loss function to replace the traditional IoU-based geometric loss metric, accelerating the model convergence process while maintaining the accuracy of bounding box regression. This allows the lightweight model to achieve practical-level detection accuracy within a limited number of iterations. Ultimately, the entire algorithm solution does not rely on high-power GPU acceleration and can complete end-to-end real-time inference at a stable frame rate on mining edge computing boxes, truly achieving organic unity and providing a complete solution for safety monitoring of underground conveyor belt transportation in coal mines that combines theoretical advancement with engineering feasibility. Attached Figure Description

[0066] Figure 1 This is the overall technical roadmap of the foreign object detection method for coal mine conveyor belts based on N-NanoDet of the present invention;

[0067] Figure 2 This is an overall structural diagram of the memory-based physical prior guidance module in this invention;

[0068] Figure 3 This is a detailed structural diagram of the physical prior memory storage module in this invention;

[0069] Figure 4 This is a detailed structural diagram of the memory-enhanced physical prior guidance module in this invention;

[0070] Figure 5 This is a detailed structural diagram of the multi-scale scene radiation recovery module in this invention;

[0071] Figure 6 This is a structural diagram of the edge constraint module based on dynamic adjustment of edge strength in this invention;

[0072] Figure 7 This is an example image of the dim light and dust synthetic dataset used for training in this invention;

[0073] Figure 8 This is a feature extraction structure diagram of the lightweight ShuffleNet_V2 convolutional neural network module in this invention;

[0074] Figure 9 This is a diagram of the N-NanoDet model network structure in this invention;

[0075] Figure 10 This is a diagram of the actual deployment module at the edge of the conveyor belt foreign object detection system in this invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] like Figures 1-10 As shown, this invention introduces a method for detecting foreign objects on coal mine conveyor belts based on N-NanoDet, comprising the following three core steps:

[0078] Step 1: Image Enhancement. Acquire the original image frames from the coal mine conveyor belt monitoring video. Enhance the original image frames based on physical priors and edge constraints to obtain an enhanced image. Specifically, the enhancement process involves encoding, compressing, and storing the prior information of historical frames in a local prior memory matrix using a physical prior memory storage module. A memory-enhanced physical prior guidance module then uses this prior memory matrix to enhance the prior features of the current frame. A multi-scale scene radiation recovery module captures the spatiotemporal correspondence between frames to restore the degraded image. Finally, an edge constraint module performs fidelity processing on the object edge details in the restored image.

[0079] Step 2: Feature Extraction and Fusion. The enhanced image is input into the lightweight feature extraction network ShuffleNet_V2 to extract multi-level foreign object features. A bidirectional multi-scale feature pyramid is constructed using FPN and PAN, and the multi-level foreign object features are fused to obtain the fused features.

[0080] Step 3: Object Detection. The fused features are input into an N-NanoDet detection network based on an anchorless bounding box mechanism. Boundary regression is performed using an improved SIoU loss function to identify and locate foreign objects in the enhanced image.

[0081] The following provides a detailed explanation of the specific implementation of each step.

[0082] The image enhancement step aims to address image degradation caused by low illumination and high dust environments in mines, providing high-quality input data for subsequent feature extraction and target detection. This step is completed collaboratively by four core modules: a physical prior memory preservation module, a memory-enhanced physical prior guidance module, a multi-scale scene radiometric restoration module, and an edge constraint module based on dynamic adjustment of edge intensity. The overall enhancement process is as follows:

[0083] First, the ConvNeXt backbone network is used to extract multi-scale input features, encoding the original degraded image into multi-scale initial scene features and initial prior features. Here, H, W, and C represent the feature's height, width, and number of channels, respectively. The initial prior features contain physical information such as transmission graph priors and atmospheric light priors, providing a priori foundation for subsequent physical model-driven enhancements.

[0084] The physical prior memory storage module is used to encode, compress, and store the physical prior information from historical video frames into a local memory matrix to form long-term memory for the current frame to access. Since underground video is long-sequence data, dust concentration and lighting conditions vary at different times; storing historical prior information provides rich reference data for enhancing the current frame.

[0085] The specific structure of the physical prior memory storage module is as follows: it employs a 3×3 convolutional kernel and a 4-layer deep backbone branch, as well as a 3×3 convolutional kernel and a 2-layer deep mask branch. Initial prior features. The features, learned through the encoder and backbone network, include prior information such as transport maps and atmospheric light. Mask branch information is enhanced by additive fusion of the backbone branch information, and multi-level features are fused through dense connections. This design enhances the representational power of the features while preserving their original spatial structure, and simultaneously outputs the key prior features for the current frame. R N×D×C This key prior feature is then processed through two pooling layers to obtain the prior compressed memory token p. R D×C Where D represents the token length, it achieves a compact compressed representation of high-dimensional prior features. The prior compressed memory token is stored in the local prior memory matrix K for subsequent cross-frame memory enhancement. The core innovation of the physical prior memory storage module lies in reducing storage overhead while maintaining the spatial structure of prior features through a dense residual feature fusion structure, making it possible to store and retrieve prior information of long-sequence videos on edge devices.

[0086] The memory-enhanced physical prior guidance module calculates the similarity weight between the current frame and the historical prior memory matrix, guiding the attention mechanism to utilize the memory information of historical frames to enhance the physical prior features of the current frame, thereby improving the ability to identify dust areas. Specifically, the similarity weight S between the key memory feature key of the current frame and the prior memory matrix K is first calculated using the following formula:

[0087] S = softmax ( )

[0088] Here, c1 serves as a normalization factor to align the dimensions of key and K. A similarity matrix S is obtained by comparing the similarity between the current frame and historical memories. The similarity weights reflect the contribution of each historical memory token to the current frame; memory tokens more similar to the current frame are assigned higher weights.

[0089] The similarity weights then guide the attention mechanism to enhance the current prior features, specifically as follows:

[0090] P = softmax( )S

[0091] Wherein, the query vector Q is obtained by using the initial prior features Expanding, parameter c2 is used as a normalization factor to align the dimensions of Q and V. After attention weighting, the memory-enhanced prior feature P is obtained. Finally, the enhanced prior feature is concatenated with the initial scene feature as input, and multiple convolutional layers are used to generate prior-guided scene features. This mechanism enables the current frame to adaptively extract effective information from historical prior memory, avoiding the failure of fixed prior assumptions in dynamic dust environments.

[0092] The multi-scale scene radiometric restoration module aims to capture the spatiotemporal dependencies across multiple spatiotemporal ranges. Through spatiotemporal feature alignment and multi-range information aggregation, it ultimately outputs clear and stable video frames.

[0093] First, features from adjacent frames are divided into multiple sets of different ranges to explore cues across different time intervals. For example, connectivity features of different ranges are processed separately to simultaneously capture spatiotemporal dependencies that change rapidly in the short term and evolve slowly in the long term. Connectivity features of different ranges are sent to a shared spatiotemporally deformable attention block, which warps the features to the target frame to achieve spatiotemporal alignment.

[0094] In a spatiotemporally deformable attention block, given the characteristics of the connection and normalized initial spacetime flow First, calculate the initial alignment feature map. :

[0095] = S( , )

[0096] Where S represents a differentiable spatiotemporal sampling operation, and then the refined spatiotemporal flow is obtained by calculating the flow offset residual. :

[0097] = ( )+

[0098] in This lightweight offset network, composed of convolutional layers, compensates for initial spatiotemporal flow estimation biases through learnable flow offset residuals, achieving more accurate inter-frame pixel-level alignment. Subsequently, cross-attention extracts temporal information from the alignment features. Used as a query for the target frame, alignment features As keys and values, range characteristics The calculation is as follows:

[0099] = W MSA(Q,K,V)

[0100] W-MSA is a multi-head self-attention mechanism for windows.

[0101] Alignment features of different ranges contain their own spatiotemporal cues and are aggregated under the guidance of prior features. Specifically, cascaded range features are treated as keys and values, multiplied by the target frame query features to calculate scene attention weights; simultaneously, prior features are treated as queries and the aligned prior features are used as attention keys to calculate prior attention weights. The final attention weight is the sum of the scene attention weight and the prior attention weight, which is then weighted and aggregated after Softmax normalization. This dual-path attention design allows multi-range feature fusion to be constrained by both data-driven and physical models, improving the clarity and stability of the output frame.

[0102] To address the problem that deviations in atmospheric light scattering model parameter estimation in non-uniform dust scenarios can easily lead to blurred or lost edge details, this invention designs an edge constraint module based on dynamic adjustment of edge intensity. This edge constraint module achieves adaptive and fidelity-preserving processing of edge details through three steps: grayscale conversion, Sobel edge detection, and weighted fusion.

[0103] Grayscale conversion step: Convert the input color image to grayscale by adding the R, G, and B channels of each pixel and then dividing by 3 to obtain the grayscale pixel value. The formula is as follows:

[0104] Gray = (R + G + B) / 3

[0105] Edge detection steps: Edge detection is performed using the Sobel operator. The Sobel operator calculates gradients in the image through convolution operations, calculating the intensity of gradient changes in the horizontal and vertical directions for each pixel in the image, thus determining the edge location and intensity. The Sobel convolution kernel is used for detecting horizontal edges. x for:

[0106] Sobel x =

[0107] Sobel convolution kernel for detecting vertical edges y for:

[0108] Sobel y =

[0109] Weighted fusion step: Calculate the edge magnitude (EM) for each pixel. Edge magnitude reflects the intensity of edges in the image; a larger gradient value indicates a more pronounced edge. The EM calculation formula is as follows:

[0110] EM =

[0111] Among them, G x and G y The horizontal and vertical gradients of the image are obtained through Sobel convolution operations, respectively; the edge constraint image is calculated using a weighted formula.

[0112] E=E output (1) EM) E raw EM

[0113] Among them, E output For grayscale images, E raw The image represents the original color image, and EM represents the normalized edge amplitude weights. Larger edge amplitudes indicate more dramatic edge changes in the region, thus assigning higher weights to the original image to preserve more detail; conversely, lower amplitudes in smoother regions result in higher weights for the grayscale image to suppress noise. The edge constraint module effectively avoids the influence of errors in atmospheric light scattering model estimation parameters, making it more suitable for complex edge variations in non-uniform dust environments underground.

[0114] The feature extraction and fusion steps aim to efficiently extract multi-level discriminative features from the enhanced image and enhance the model's ability to perceive foreign objects of different sizes through multi-scale fusion. Specifically, this includes the following three steps:

[0115] (1) Lightweight feature extraction based on ShuffleNet_V2

[0116] ShuffleNet_V2 is used as the backbone network. This backbone network is designed for devices with limited computing resources and has the advantages of few parameters, low computational cost, fast processing speed, and high recognition accuracy. The input image first extracts basic features through a 3×3 convolutional layer, followed by max pooling to reduce the feature map size and enhance the abstraction of the features. Next, three multi-level feature extractions are performed, each time reducing the feature map size through downsampling and increasing the receptive field to extract deeper information.

[0117] In the downsampling operation, the input feature map is divided into channels, with both the left and right branches having the same number of channels, half the original number. The left branch does not perform convolution operations to reduce computation, while the right branch contains three consecutive convolutions: two ordinary convolutions and one depthwise separable convolution. This design reduces computational complexity while preserving effective feature extraction capabilities. Subsequently, the outputs of the left and right branches are merged using the Concat method, and information exchange is achieved using Channel Shuffle. Through this mechanism, ShuffleNet_V2 extracts three layers of discriminative features, including information on foreign object edges, shape, and location, providing a rich feature foundation for subsequent multi-scale fusion.

[0118] (2) Multi-scale feature fusion based on FPN and PAN

[0119] FPN generates multi-scale feature maps through a top-down structure and fuses information from different scales at each level of the pyramid. First, feature maps of different scales output from each stage of ShuffleNet_V2 are used as input. Deeper, low-resolution feature maps are upsampled to align with higher-resolution feature maps, allowing information to flow between different scales. The fusion formula is:

[0120] P l = UpSample(C l )+C l+1

[0121] Among them, P l Features of the first layer of the pyramid, C l For the corresponding layer features of the backbone network, UpSample is the upsampling operation. Through the FPN structure, deep semantic information is passed to the shallow layer, which enhances the classification and discrimination ability of the shallow features.

[0122] PAN further enhances the performance of FPN by using bottom-up path aggregation to bring shallow location information back to deeper layers. The path aggregation operation formula is as follows:

[0123] P agg = Conv2D(P l )

[0124] Among them, P agg P represents the features after path aggregation. l These are feature maps extracted at different levels. PAN aggregates shallow and deep features to facilitate smoother information flow, ensuring a good fusion of precise localization information from the shallow layers and rich semantic information from the deep layers.

[0125] (3) Bidirectional feature fusion and weighted aggregation

[0126] This system combines the efficient feature extraction capabilities of ShuffleNet_V2 with the multi-scale feature fusion capabilities of FPN and PAN to form a bidirectional feature fusion system. Deep semantic information is processed through the feature pyramid of FPN and then fused with shallow location information through path aggregation in PAN to form a multi-scale, rich feature representation. Finally, a weighted fusion mechanism is used to match the weights of features at different scales to highlight foreign object features. The weighted fusion formula is as follows:

[0127] F final =

[0128] in, The weights of the features at each scale are... These are feature maps at different scales. Through the weighted fusion described above, the feature responses of important targets are enhanced, irrelevant background information is effectively suppressed, and finally, a multi-scale fused feature map with both strong semantics and high localization quality is obtained.

[0129] In this invention, the target detection step is based on the N-NanoDet lightweight detection network, and uses an anchorless mechanism and an improved SIoU loss function to accurately identify and locate foreign objects in the fused feature map.

[0130] (1) Anchorless detection mechanism

[0131] The N-NanoDet model employs an anchor-free mechanism to eliminate the dependence on predefined anchor boxes. Traditional anchor-box methods perform target detection by predefining boxes of different scales and aspect ratios, which, while effective, significantly increases computational and memory consumption. This invention uses an anchor-free method to directly predict the thermal response value of the center point of the foreign object on the belt, the center point offset, the width and height of the bounding box, and the target class confidence from each spatial location in the feature map through a convolutional network. This avoids the sensitive impact of manually designed anchor box hyperparameters on detection performance, while reducing redundant computation and improving inference speed.

[0132] (2) Lightweight network design

[0133] The N-NanoDet model extensively employs depthwise separable convolutions, decomposing standard convolutions into channel-wise spatial convolutions and pointwise 1×1 convolutions, thus compressing the number of model parameters to 1 / 8 to 1 / 6 of that of traditional convolutions. Simultaneously, through feature concatenation operations, the fused feature maps of different scales output by PAN are input into the corresponding Head sub-networks for joint discrimination. Each Head sub-network shares convolution parameters, and finally outputs the bounding box coordinates and class confidence of each detected target. This design maintains high detection accuracy while enabling the model to run efficiently on low-power edge devices.

[0134] (3) Improved SIoU loss function

[0135] This invention employs an improved SIoU loss function for bounding box regression, and the total loss LSIoU is expressed as:

[0136] L SIoU =1 IoU+

[0137] Wherein, IoU is the overlap area loss, which measures the degree of overlap between the predicted box and the ground truth box; To compensate for the distance loss, the offset of the object's position was calculated and the effect of the angle on the position was corrected. An angle penalty factor was introduced to compensate for the impact of the position offset on the regression accuracy. The aspect ratio difference loss measures the difference in the aspect ratio of the target bounding box. The joint optimization of the triple loss function enables the model to achieve accurate target localization and precise bounding box regression even in low-quality downhole images.

[0138] To address the challenge of obtaining large quantities of realistic labeled data from underground coal mines, this invention also provides a method for synthesizing training datasets for coal mine dust environments, used to construct physically realistic training datasets. Specifically, a DPTHybrid depth estimation model is used to predict the depth of each frame of the input image pixel by pixel to obtain detailed scene geometric structure information; subsequently, a spatially distributed dust concentration density map is constructed through depth inversion, thereby simulating the absorption and scattering effects of particulate matter on incident light at different distance levels.

[0139] A fogging synthesis is performed on a sharp image based on an atmospheric scattering model. The mathematical expression of the atmospheric scattering model is:

[0140] I(x) = J(x)·t(x) + A·(1 t(x))

[0141] Where I(x) is the foggy image, J(x) is the clear, fog-free image, t(x) is the transmittance, and A is atmospheric light. To suit coal mine dust environments, the transmittance is dynamically adjusted based on dust concentration, and the calculation formula is as follows:

[0142] t(x) = exp(-β·d(x)·C(x))

[0143] Where d(x) is the distance to the light source, C(x) is the dust concentration at that location, and β is the dust attenuation coefficient, representing the dust's absorption and scattering effect on light.

[0144] To mitigate the impact of depth estimation noise on transmittance continuity, a joint optimization mechanism combining bilateral filtering and guided filtering is introduced to achieve a smooth transition in the transmittance map while preserving edge details. Next, an adaptive estimation of atmospheric light A(x) is performed using a dark channel prior method, and the illumination distribution is dynamically corrected based on local brightness and dust concentration variations, making the generated fogged image closer to the real-world downhole scene in terms of brightness and contrast. Furthermore, to enhance the diversity and physical reliability of the synthesized data, fogged image samples with low, medium, and high fog density levels are synthesized for each frame, effectively covering multi-scale visibility variations in the actual environment.

[0145] This invention also provides a system solution for deploying the above-mentioned detection method as an edge device in actual industrial applications. The system includes an industrial camera, an edge computing box, an alarm, a microswitch, and a monitoring platform. The industrial camera is installed above the feed chute of a belt conveyor. The camera's exposure and gain parameters are automatically adjusted according to the ambient light level in the mine, used to acquire video images of the conveyor surface in real time and output them as an RTSP video stream. The edge computing box uses an ARM architecture processor and is pre-installed with an Android operating system. It carries a quantized and compressed model weight file and a lightweight inference engine to run the image enhancement, feature extraction, and target detection operations. The low power consumption and high energy efficiency of the ARM architecture processor are highly compatible with the industrial requirements of explosion-proof and energy-saving mining equipment. The edge computing box runs the complete detection algorithm locally, realizing intelligent processing at the detection front end and avoiding the network bandwidth pressure and transmission latency caused by transmitting large amounts of video data to the cloud for processing.

[0146] When the edge computing box detects a foreign object on the conveyor belt surface, it controls the alarm to issue a tiered alarm via a microswitch. Simultaneously, it uploads and saves the processing results and image data to the monitoring platform. The monitoring platform communicates with the edge computing box via industrial Ethernet, receiving and storing detection results, alarm records, and keyframe image data before and after the alarm in real time, and providing a time-axis-based backtracking retrieval interface. This system forms an integrated closed-loop control system of "local real-time detection—front-end instant alarm—remote data management," providing a complete solution for safety monitoring of underground conveyor belt transportation in coal mines, combining real-time performance, reliability, and engineering feasibility.

[0147] In summary, this invention addresses the triple challenges of foreign object detection on conveyor belts in low-light, high-dust environments of underground coal mines: environmental degradation, difficulty in perceiving small targets, and limited edge computing power. It proposes a lightweight detection method based on N-NanoDet. In the image enhancement stage, a physical prior memory preservation and memory enhancement guidance mechanism utilizes prior information from historical frames to recover degraded images, and an edge constraint module is introduced to preserve object edge details. In the feature extraction stage, a bidirectional multi-scale feature pyramid is constructed based on ShuffleNet_V2 combined with FPN and PAN to achieve deep interactive fusion of shallow localization and deep semantic features. In the detection stage, an anchor-free mechanism and depthwise separable convolution are used to achieve model lightweighting, while an improved SIoU loss function ensures detection accuracy. Furthermore, a method for synthesizing physically realistic datasets for dusty environments and an edge device deployment scheme are provided. Experiments show that this invention outperforms existing methods in terms of detection accuracy, inference speed, and edge deployment efficiency, providing a solution for coal mine conveyor belt safety monitoring that combines theoretical advancement with engineering practicality.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet, characterized in that, include: 1) Image enhancement step: Obtain the original image frames of the coal mine conveyor belt monitoring video, and perform enhancement processing on the original image frames based on physical priors and edge constraints to obtain an enhanced image. The enhancement processing encodes and compresses the prior information of historical frames and stores it in the local prior memory matrix through the physical prior memory storage module. The memory enhancement physical prior guidance module uses the prior memory matrix to enhance the prior features of the current frame. Then, the multi-scale scene radiation recovery module captures the spatiotemporal correspondence between frames to restore the degraded image. Finally, the edge constraint module performs fidelity processing on the object edge details of the restored image. 2) Feature extraction and fusion step: The enhanced image is input into the lightweight feature extraction network ShuffleNet_V2 to extract multi-level foreign object features, and a bidirectional multi-scale feature pyramid is constructed using FPN and PAN. The multi-level foreign object features are then fused to obtain the fused features; 3) Target detection step: The fused features are input into the N-NanoDet detection network based on the anchorless frame mechanism, and regression is performed through the loss function to identify and locate foreign objects in the enhanced image.

2. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 1, characterized in that, The image enhancement steps specifically include: 1) Use the ConvNeXt backbone network to extract multi-scale input features, and encode and compress the transmission map and atmospheric light prior information into prior memory tokens and save them to the local prior memory matrix through the dense residual feature fusion structure in the physical prior memory storage module. 2) The similarity weight between the current frame and the historical prior memory matrix is ​​calculated by the memory-enhanced physical prior guidance module, which guides the attention mechanism to generate prior-guided scene features. The spatiotemporal deformable attention block of the multi-scale scene radiation recovery module is used to capture the spatiotemporal correspondence between frames, so as to output a clear image with dehazing and blurring. 3) The color image is converted to grayscale using an edge constraint module based on dynamic adjustment of edge intensity. Then, the Sobel operator is used to calculate the gradients in the horizontal and vertical directions to obtain the edge amplitude of each pixel. Finally, the image is weighted and fused based on the edge amplitude to preserve the edge details of the object.

3. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 2, characterized in that, The physical prior memory storage module employs a 3×3 convolutional kernel and a 4-layer deep backbone branch, as well as a 3×3 convolutional kernel and a 2-layer deep mask branch, for the initial prior features. The information is learned through the encoder and backbone network, including prior information such as transmission maps and atmospheric light, where H, W, and C represent the feature height, width, and number of channels, respectively. The mask branch information enhances the initial prior feature information through additive fusion and fuses multi-level features through dense connections to output the key prior feature of the current frame. R N×D×C The key prior features are processed by two pooling layers to obtain a prior compressed memory token p. R D×C Where D represents the token length, and the prior compressed memory token is stored in the local memory matrix.

4. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 2, characterized in that, In the memory-enhanced physical prior guidance module, the similarity weight S between the key memory feature key of the current frame and the prior memory matrix K is calculated, specifically as follows: S =softmax ( ), Here, c1 serves as a normalization factor to align the dimensions of key and K. A similarity matrix is ​​obtained by comparing the similarity between the current frame and the memory. The similarity weights then guide the attention mechanism to enhance the current prior features, specifically as follows: P=softmax( )S, Wherein, the query vector Q is obtained by using the initial prior features Expanding, parameter c2 is used as a normalization factor to align the dimensions of Q and K, ultimately obtaining the memory-enhanced prior feature P. The enhanced prior feature is concatenated with the initial scene feature as input, and scene features guided by the prior are generated through multiple convolutional layers.

5. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 2, characterized in that, The multi-scale scene radiometric reconstruction module divides adjacent frame features into multiple sets with different ranges to explore clues at different time intervals. Connectivity features of different ranges are sent to a shared spatiotemporally deformable attention block. Within the spatiotemporally deformable attention block, features of a given connection are... and normalized initial spacetime flow First, calculate the initial alignment feature map. : = S( ， ), Where S represents a differentiable spatiotemporal sampling operation; then, the refined spatiotemporal flow is obtained by calculating the flow offset residual. : , in It is a lightweight offset network composed of convolutional layers; cross attention extracts temporal information from alignment features. Used as a query for the target frame, alignment features As keys and values, range characteristics The calculation is as follows: = W MSA(Q,K,V), Alignment features of different ranges are aggregated under the guidance of prior features. The concatenated range features are regarded as keys and values, multiplied by the target frame query features to calculate scene attention weights. The prior features are regarded as queries and the aligned prior features are used as attention keys to calculate prior attention weights. The final attention weight is the sum of scene attention weights and prior attention weights. After Softmax normalization, the range features are weighted and aggregated.

6. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 2, characterized in that, In the edge constraint module based on dynamic adjustment of edge intensity, the input color image is converted to grayscale, and the R, G, and B channels of each pixel are added together and then divided by 3 to obtain the grayscale pixel value. Edge detection is computed using the Sobel operator; the Sobel convolution kernel is used to detect horizontal edges. x for: Sobel x = , Sobel convolution kernel for detecting vertical edges y for: Sobel y = , The formula for calculating the edge amplitude EM of each pixel is: EM = , Among them, G x and G y The horizontal and vertical gradients of the image are obtained through convolution operations, respectively; the edge constraint image is calculated using a weighted formula. E=E output (1 IN) AND raw IN, Among them, E output For grayscale images, E raw This is the original color image.

7. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 1, characterized in that, The feature extraction and fusion steps specifically include: The ShuffleNet_V2 network serves as the backbone network. Through a channel partitioning mechanism, the input feature map channels are divided into left and right branches. The left branch does not undergo convolution, while the right branch performs downsampling through continuous convolution and depthwise separable convolution. Subsequently, the output feature maps of the two branches are combined through channel merging and channel shuffling mechanisms to extract the three-layer features of the object's edge, shape, and location layer by layer. The extracted multi-layer features are input into the FPN, where a top-down upsampling operation aligns and fuses the deep low-resolution features with the shallow high-resolution features. The fusion formula is as follows: P l = UpSample(C l )+C l+1, Among them, P l For the features of the first layer of the pyramid, C l For the corresponding layer features of the backbone network, UpSample is the upsampling operation; the features processed by FPN flow to PAN, and the shallow layer's location information flows back to the deep layer through bottom-up path aggregation. The path aggregation operation formula is: P agg = Conv2D(P l ), Among them, P agg P represents the features after path aggregation. l Feature maps are extracted at different levels; finally, a weighted fusion mechanism is used to match the weights of features at different scales to highlight foreign object features. The weighted fusion formula is as follows: F final = , in, The weights of the features at each scale are... By generating feature maps at different scales, the final multi-scale fused feature map is obtained.

8. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to claim 1, characterized in that, The target detection steps specifically include: The N-NanoDet model employs an anchor-free mechanism to eliminate the dependence on predefined anchor boxes, and uses a convolutional network to directly perform position regression prediction on the centrality, width, height, and category of foreign objects on the belt. Depthwise separable convolutions are applied to the model to reduce the computational cost and number of parameters. Features at different levels are input into the Head module through feature concatenation operations, and finally the bounding boxes of each detected target are output. The loss function is an improved loss function, with a total loss L. SIoU Represented as: L SIoU =1 IoU+ , in, For the overlapping area loss, To account for distance loss, the object's position offset was calculated and the effect of angle on position was corrected. The aspect ratio difference loss measures the aspect ratio difference of the target box, thereby improving the prediction accuracy of the target box shape.

9. The method for detecting foreign objects in coal mine conveyor belts based on N-NanoDet according to any one of claims 1 to 8, characterized in that, It also includes a dataset synthesis step for building the training dataset: The DPT Hybrid depth estimation model is used to perform pixel-by-pixel depth prediction for each frame of input image to obtain scene geometric structure information, and a spatially distributed dust concentration density map is constructed through depth inversion. According to the atmospheric scattering model I(x) = J(x)·t(x) + A·(1 t(x)) is used to perform fogging synthesis on a sharp image, where J(x) is the sharp image, t(x) is the transmittance, and A is the atmospheric light; The transmittance t(x) is dynamically adjusted according to the dust concentration, t(x) = exp(-β·d(x)·C(x)), where d(x) is the distance to the light source, C(x) is the dust concentration at that location, and β is the dust attenuation coefficient; A joint optimization mechanism of bilateral filtering and guided filtering is introduced to smooth the transmission map. The atmospheric light A(x) is adaptively estimated using the dark channel prior method, and the illumination distribution is dynamically corrected by combining local brightness and dust concentration changes. For each frame, three fog image samples with different fog density levels of low, medium and high are synthesized to cover the multi-scale visibility changes in the actual environment.

10. A foreign object detection system for coal mine conveyor belts based on N-NanoDet, employing the method as described in any one of claims 1 to 9, characterized in that, The system includes an industrial camera, an edge computing box, an alarm, a micro switch, and a monitoring platform. The industrial camera is installed above the feed chute of the belt conveyor to acquire real-time images of the belt. The edge computing box uses an ARM architecture processor and is pre-installed with an Android operating system to process the real-time images. When the edge computing box detects a foreign object on the surface of the belt, it controls the alarm to sound an alarm via the micro switch, and simultaneously uploads and saves the processing results and image data to the monitoring platform.

Images