A method for intelligent visual recognition and sorting control of large mining materials

By using an improved YOLOv8n model to identify large materials in the underground environment, and combining adaptive feature enhancement and image inpainting techniques, the problem of image quality degradation in underground visual recognition systems was solved, achieving efficient and accurate material detection and positioning.

CN123102647APending Publication Date: 2026-11-17张家港市沙洲智能设备制造有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611308522.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to stably and accurately detect and locate large materials in low-light, high-dust, and high-vibration environments underground. The image quality of traditional visual recognition systems drops sharply under these conditions, causing target detection algorithms to fail.

Method used

An improved YOLOv8n model with multi-source parameter synchronous input is adopted, combined with an adaptive pre-feature enhancement branch, a blind dehazing and deblurring fusion unit, and a spatiotemporal linkage detection and tracking head to suppress reflection interference, enhance dark textures, and achieve end-to-end image inpainting and target tracking.

Benefits of technology

It significantly improves the accuracy and real-time performance of material identification under extreme working conditions, reduces computing power consumption, and meets the needs of intelligent sorting in underground mines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123102647A_ABST
    Figure CN123102647A_ABST
Patent Text Reader

Abstract

This application provides an intelligent visual recognition and sorting control method for large mining materials. The method includes: S1: acquiring an original image of the large ore material and simultaneously collecting supplementary parameters corresponding to the original image; S2: inputting the original image and the supplementary parameters into an improved YOLOv8n model and outputting sorting control commands. The improvements include: setting an adaptive pre-feature enhancement branch between the image input layer and the backbone network; setting a lightweight blind dehazing and deblurring fusion unit in the neck network; and integrating a spatiotemporal linkage detection and tracking head for velocity prediction and target matching; S3: completing an adaptive sorting operation based on the sorting control commands. This method, through synchronous input of multi-source parameters and collaboration with the improved YOLOv8n model, effectively suppresses reflective interference and enhances dark textures in low-light, high-dust, and strong-vibration environments, dynamically repairs image degradation, improves recognition accuracy, and meets the needs of intelligent sorting in underground mining.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of mineral sorting technology, and in particular to an intelligent visual recognition and sorting control method for large mineral materials. Background Technology

[0002] In the field of mineral sorting technology, especially for the intelligent identification and sorting of large materials underground, existing technologies mainly rely on manual picking or traditional mechanical screening methods. However, with the increase in the depth of mineral resource mining and the advancement of intelligent mine construction, the underground working environment poses extremely stringent challenges to visual recognition systems.

[0003] Currently, while some research attempts to introduce machine vision technology into mineral sorting, most are limited to surface coal preparation plants or scenarios with relatively stable lighting conditions. In actual underground operations, mining faces are typically in low-light environments, with material details submerged in darkness. Local lighting, such as miner's lamps, easily creates specular reflections on the ore surface, resulting in extreme contrast phenomena where images exhibit both "dead black" and "overexposed" colors. Simultaneously, the dust concentration in crushing and transport processes is extremely high. Suspended particles not only scatter light, causing image blurring, but also continuously adhere to the lens surface, causing a rapid decline in image quality within a short period. Furthermore, the continuous vibration of heavy equipment and the increased speed of conveyor belts further contribute to dynamic motion blur and jitter in the images. The combined effect of these multiple environmental factors makes it difficult for traditional target detection algorithms to operate stably underground: conventional image enhancement methods cannot effectively recover texture information in reflective areas; single-modality vision sensors struggle to penetrate dust interference; and general detection and tracking models often suffer from target loss and frequent identity switching when faced with high-speed motion and vibration blurring.

[0004] Therefore, how to enable visual recognition systems to stably and accurately detect and locate large materials under extreme working conditions such as low illumination, high dust, and strong vibration has become a key technical challenge restricting the development of intelligent sorting technology in underground mines. Summary of the Invention

[0005] This application provides an intelligent visual recognition and sorting control method for large mining materials. This method, through synchronous input of multi-source parameters and collaboration with an improved YOLOv8n model, effectively suppresses reflective interference and enhances dark textures in low-light, high-dust, and strong-vibration environments, dynamically repairs image degradation, realizes end-to-end speed prediction and tracking, significantly improves recognition accuracy and real-time performance, and reduces computing power overhead, thus meeting the needs of intelligent sorting in underground mines.

[0006] In a first aspect, a method for intelligent visual recognition and sorting control of large mining materials is provided, the method comprising: S1: Acquire the original image of the large ore material, and simultaneously collect the supplementary parameters corresponding to the original image. The supplementary parameters include: imaging exposure supplementary lighting parameters, field of view environment parameters, pixel physical space coordinate transformation matrix, and real-time operation parameters of the conveyor belt. S2: Input the original image and the supplementary parameters into the improved YOLOv8n model and output sorting control instructions. The improvement measures include: setting an adaptive pre-feature enhancement branch between the image input layer and the backbone network, setting a lightweight blind dehazing and deblurring fusion unit in the neck network, and integrating a spatiotemporal linkage detection and tracking head for velocity prediction and target matching. S3: Perform adaptive sorting operation based on the sorting control command.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the pre-feature enhancement branch receives the imaging exposure supplementary lighting parameters, divides the input original image into regions, identifies and filters specular reflection regions without effective texture, performs adaptive dynamic range expansion processing on the remaining effective regions, performs directional texture feature enhancement on the dark regions, and outputs a fused feature map to the backbone feature extraction network.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the pre-feature enhancement branch includes a reflection suppression submodule and a dark area enhancement submodule; The reflection suppression submodule is used to identify specular reflection areas, generate an invalid feature marker matrix, and suppress the feature weights of specular reflection areas. The dark area enhancement submodule extracts the lateral and longitudinal gradient features of the dark area of ​​large ore materials. It strengthens the strong gradient features corresponding to ore joints and suppresses the weak gradient features corresponding to homogeneous ore areas through learnable convolution weighting. The weighted gradient features and the grayscale features after dynamic range expansion are then concatenated and output.

[0009] It should be understood that this pre-feature enhancement branch, through regionalized image processing linked to imaging exposure and illumination parameters, combined with the directional optimization design of the reflection suppression submodule and the dark area enhancement submodule, solves the pain points of target contour fragmentation, false alarms, and noise amplification caused by specular reflections caused by water film and wet mud on the surface of large ore blocks in low-light, high-light illumination scenarios underground, as well as the annihilation of ore texture features in dark areas and the blurring of the classification boundary between valuable ore and waste rock. At the same time, it abandons the serial architecture of the preprocessing and detection networks in traditional vision solutions, and achieves end-to-end joint optimization with the backbone feature extraction network. While effectively enhancing the effective features of ore and improving the accuracy of target recognition and positioning, it significantly reduces the ineffective computing power overhead, adapting to the low computing power constraints of underground edge computing devices and the real-time detection requirements under extreme working conditions.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the blind dehazing and deblurring fusion unit receives the field of view environment parameters and performs defect repair on the multi-scale feature map output by the backbone feature extraction network.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the blind dehazing and deblurring fusion unit includes a local fog concentration estimation subunit, a local motion blur degree estimation subunit, an adaptive weight allocation subunit, a dehazing branch, a deblurring branch, and a bypass through branch; The local fog concentration estimation subunit and the local motion blur estimation subunit are set in parallel, and their outputs are connected to the adaptive weight allocation subunit. The outputs of the adaptive weight allocation subunit are connected to the defogging branch, the deblurring branch, and the bypass direct branch, respectively. The adaptive weight allocation subunit allocates computational weights to the defogging branch and the deblurring branch of the corresponding region according to the fog concentration and motion blur estimation results of each local region of the feature map. For regions with significant defects, the corresponding branch is used to complete feature repair, and for regions without defects, the output is directly output through the bypass direct branch.

[0012] It should be understood that this application, by setting a blind dehazing and deblurring fusion unit that can receive field-of-view environmental parameters, can complete the adaptation defect repair for the multi-scale feature map output by the backbone feature extraction network. While effectively solving the problem of the separation between dehazing and deblurring processing in existing visual processing technologies and the difficulty in adapting to the non-uniform degradation of the field of view, it improves the robustness and environmental adaptability of the visual system in complex and harsh field-of-view environments.

[0013] By using parallel local fog concentration estimation subunits and local motion blur estimation subunits, the system can accurately perceive the degradation type and degree of each local region in the feature map, providing a precise basis for subsequent regionalization processing. Through an architecture that combines adaptive weight allocation subunits with defogging branches, deblurring branches, and bypass direct access branches, the system can assign appropriate computational weights to the corresponding defogging and deblurring branches based on the fog concentration and motion blur estimation results of each local region. This allows for targeted feature repair by using the corresponding branches in areas with significant defects, while also enabling direct output of the original features from defect-free areas through bypass direct access branches. This effectively prevents information loss in defect-free areas due to unnecessary processing, ensuring the integrity of the feature map and the retention of original effective information.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, the spatiotemporal linkage detection and tracking head adds a lightweight target velocity prediction branch and a lightweight target matching unit to the original classification branch, localization branch, and confidence branch of YOLOv8n; the velocity prediction branch shares underlying features with each branch, and its output is connected to the target matching unit.

[0015] In conjunction with the first aspect, in some implementations of the first aspect, the target speed prediction branch receives the pixel physical space coordinate transformation matrix and the real-time operating parameters of the conveyor belt, predicts the motion trajectory of the material target, the position and time of arrival at the sorting execution mechanism, and compensates for the model inference delay and the execution mechanism response delay; the lightweight target matching unit assigns a unique tracking identifier to each material target to achieve continuous trajectory matching between frames of targets, and avoids target misalignment and repeated detection under high-speed conveying.

[0016] It should be understood that the spatiotemporal linkage detection and tracking head achieves end-to-end integration of detection and tracking. Based on the native classification, localization, and confidence branches, the velocity prediction branch, which shares underlying features, can receive the pixel spatial transformation matrix and conveyor belt operation parameters in real time, accurately predicting the material's trajectory and the time to reach the actuator, effectively compensating for the delay caused by model inference and mechanical response. At the same time, the lightweight target matching unit assigns a unique tracking identifier to each material, ensuring continuous and stable target matching between frames under high-speed conveying, avoiding target misalignment and repeated detection problems, and significantly improving the accuracy of dynamic grasping and system response efficiency.

[0017] In conjunction with the first aspect, in some implementations of the first aspect, the improved YOLOv8n lightweight target detection model compares the action execution data fed back from the sorting execution stage with the detection results and sorting prediction results output by the model to identify and collect false detection and missed detection samples. After completing the automatic labeling of samples, incremental training is performed, and the model weights are updated online during the inference interval. Attached Figure Description

[0018] Figure 1 The flowchart illustrates the implementation of an intelligent visual recognition and sorting control method for large mining materials provided in this application embodiment. Detailed Implementation

[0019] The terminology used in the following embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to also include expressions such as “one or more,” unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of this application, “at least one” and “one or more” refer to one, two, or more than two. The term “and / or” is used to describe the relationship between related objects, indicating that three relationships can exist; for example, A and / or B can indicate: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character “ / ” generally indicates that the preceding and following related objects are in an “or” relationship.

[0020] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0021] In the field of mineral sorting, the identification of large materials underground has long relied on manual or mechanical screening, which is inefficient. With the advancement of intelligent mine construction, extreme underground working conditions pose severe challenges to visual recognition: the illumination at the mining face is extremely low, the details in the dark areas of the material are lost, and the miners' lamps create a mirror-like reflection on the ore surface, resulting in images that are both "completely black" and "overexposed"; the dust concentration in the crushing and transfer process is very high, causing image blurring and lens contamination; the vibration of heavy equipment and the high-speed operation of conveyor belts further cause dynamic motion blur. These factors make it difficult for traditional algorithms to work stably, conventional enhancement methods cannot recover reflective textures, single sensors cannot penetrate dust, and general tracking models are prone to target loss.

[0022] Therefore, how to achieve stable material identification in underground environments with low illumination, high dust, and strong vibration has become a key technical challenge restricting the development of intelligent sorting.

[0023] This application provides an intelligent visual recognition and sorting control method for large mining materials, which can effectively overcome the above-mentioned problems.

[0024] The technical solutions provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0025] Figure 1 The flowchart illustrates the implementation of an intelligent visual recognition and sorting control method for large mining materials provided in this application embodiment.

[0026] refer to Figure 1 In some examples, the method includes: S1: Acquire the original image of the large ore material, and simultaneously collect the supplementary parameters corresponding to the original image. The supplementary parameters include: imaging exposure supplementary lighting parameters, field of view environment parameters, pixel physical space coordinate transformation matrix, and real-time operation parameters of the conveyor belt. S2: Input the original image and the supplementary parameters into the improved YOLOv8n model and output sorting control instructions. The improvement measures include: setting an adaptive pre-feature enhancement branch between the image input layer and the backbone network, setting a lightweight blind dehazing and deblurring fusion unit in the neck network, and integrating a spatiotemporal linkage detection and tracking head for velocity prediction and target matching. S3: Perform adaptive sorting operation based on the sorting control command.

[0027] In some examples, the pre-feature enhancement branch receives the imaging exposure supplementary lighting parameters, divides the input raw image into regions, identifies and filters specular reflection regions without effective texture, performs adaptive dynamic range expansion processing on the remaining effective regions, performs directional texture feature enhancement on the dark regions, and outputs a fused feature map to the backbone feature extraction network.

[0028] In some examples, the front feature enhancement branch includes a reflection suppression submodule and a dark area enhancement submodule; The reflection suppression submodule is used to identify specular reflection areas, generate an invalid feature marker matrix, and suppress the feature weights of specular reflection areas. The dark area enhancement submodule extracts the lateral and longitudinal gradient features of the dark area of ​​large ore materials. It strengthens the strong gradient features corresponding to ore joints and suppresses the weak gradient features corresponding to homogeneous ore areas through learnable convolution weighting. The weighted gradient features and the grayscale features after dynamic range expansion are then concatenated and output.

[0029] In one possible implementation, the reflection suppression submodule is first executed to convert the original RGB image into a single-channel grayscale image. The grayscale threshold is dynamically adjusted based on the global average brightness in the imaging exposure lighting parameters (the higher the global average brightness, the higher the threshold is). The grayscale image is then binary segmented to obtain an initial high grayscale region mask. Simultaneously, a 5×5 sliding window is used to traverse the grayscale image, and the Sobel gradient variance within each window is calculated. Regions with gradient variance less than a preset threshold are marked as textureless regions. The initial high grayscale region mask and the textureless region mask are then logically ANDed to generate the final specular reflection binary mask. The binary mask is then expanded into an invalid feature label matrix of the same dimension as the original image (the channel value corresponding to the specular reflection region is set to 0, and the other regions are set to 1).

[0030] The shadow enhancement submodule, based on the regional illumination brightness coefficient in the imaging exposure illumination parameters, classifies areas in the original image where the illumination brightness coefficient is lower than a preset value as shadow regions. Adaptive Histogram Equalization (CLAHE) is applied to these shadow regions for dynamic range expansion, where the clipLimit parameter of CLAHE is dynamically adjusted according to the global average brightness. The Sobel operator is used to extract the horizontal and vertical gradient features of the expanded grayscale image of the shadow regions, resulting in two gradient maps of size H×W×1. These two gradient maps are then concatenated into an H×W×2 gradient feature map. This gradient feature map is weighted using two consecutive 1×1 learnable convolutional layers (16 and 16 output channels respectively), strengthening strong gradient features corresponding to ore joints and suppressing weak gradient features corresponding to homogeneous ore regions. Finally, the weighted gradient feature map is concatenated with the dynamically expanded RGB image of the shadow region to obtain the enhanced local feature map.

[0031] Finally, the original image weighted by the invalid feature label matrix and the local feature map enhanced by dark areas are fused together. The number of channels of the fused feature map is adjusted to 32 through a 1×1 convolutional layer and output to the backbone feature extraction network. At the same time, this pre-feature enhancement branch is jointly trained end-to-end with the backbone network and the detection head. During the training process, the generation logic of the invalid feature label matrix and the weights of the 1×1 convolutional layer are all involved in gradient update.

[0032] In some examples, the blind dehazing and deblurring fusion unit receives the field-of-view environment parameters and performs defect repair on the multi-scale feature map output by the backbone feature extraction network.

[0033] In some examples, the blind dehazing and deblurring fusion unit includes a local fog concentration estimation subunit, a local motion blur estimation subunit, an adaptive weight allocation subunit, a dehazing branch, a deblurring branch, and a bypass through branch; The local fog concentration estimation subunit and the local motion blur estimation subunit are set in parallel, and their outputs are connected to the adaptive weight allocation subunit. The outputs of the adaptive weight allocation subunit are connected to the defogging branch, the deblurring branch, and the bypass direct branch, respectively. The adaptive weight allocation subunit allocates computational weights to the defogging branch and the deblurring branch of the corresponding region according to the fog concentration and motion blur estimation results of each local region of the feature map. For regions with significant defects, the corresponding branch is used to complete feature repair, and for regions without defects, the output is directly output through the bypass direct branch.

[0034] In one possible implementation, the blind dehazing and deblurring fusion unit takes the P3, P4, and P5 multi-scale feature maps output from the YOLOv8n backbone network, along with synchronously acquired field-of-view environmental parameters (viewpoint pollution level, field-of-view dust concentration), as input. It processes each scale feature map independently and outputs a repaired multi-scale feature map of the same dimension. During processing, two-way defect estimation is performed in parallel on the single-scale input feature map. The local fog concentration estimation subunit extracts the dark channel features of the feature map through two consecutive 3×3 depthwise separable convolutions. After concatenation with the field-of-view environmental parameter channels, it is compressed by a 1×1 convolution and mapped using the sigmoid function to generate a region-wise fog concentration weight map in the [0,1] interval. The local motion blur estimation subunit simultaneously calculates the local gradient magnitude of the feature map, calculates the gradient variance through a sliding window, and normalizes it to generate a [0,1] interval. The region-by-region blur weight map is generated, and then the two weight maps are input into the adaptive weight allocation subunit. According to the preset rules, the region-by-region operation weights of three parallel branches, namely dehazing, deblurring, and bypass pass-through, are generated. For a single defective region, the corresponding repair branch weight is reset to 1 and the weights of the other branches are set to 0. For a defect-free region, the pass-through branch weight is reset to 1 and the repair branch weight is set to 0. For mixed defective regions, the weights are allocated by linear interpolation to ensure that the sum of the weights of the three branches is 1. Then, the input feature map is simultaneously input into the three parallel branches to complete the corresponding processing. The dehazing branch uses a lightweight dark channel prior network with depthwise separable convolution and edge-aware smoothing layers to complete feature dehazing. The deblurring branch uses a lightweight residual network with a learnable 3×3 motion blur kernel to complete feature deblurring. The bypass pass-through branch directly outputs the original feature map. Finally, the output feature maps of the three branches are fused with the corresponding region-by-region weights pixel-by-pixel to obtain a single-scale repair feature map. After completing the above processing on all scale feature maps, the repaired multi-scale feature map is output to the subsequent neck module.

[0035] In some examples, the spatiotemporal linkage detection and tracking head adds a lightweight target velocity prediction branch and a lightweight target matching unit to the original classification branch, localization branch, and confidence branch of YOLOv8n; the velocity prediction branch shares the underlying features with each branch, and its output is connected to the target matching unit.

[0036] In some examples, the target speed prediction branch receives the pixel physical space coordinate transformation matrix and the real-time operating parameters of the conveyor belt to predict the motion trajectory of the material target, the position and time of arrival at the sorting actuator, and compensate for the model inference delay and the actuator response delay; the lightweight target matching unit assigns a unique tracking identifier to each material target to achieve continuous trajectory matching between frames and avoid target misalignment and repeated detection under high-speed conveying.

[0037] In one possible implementation, the spatiotemporal linkage detection and tracking head uses the multi-scale feature map output by the YOLOv8n neck feature fusion network as a common source shared input. After extracting common features through the bottom shared convolutional layer of the detection head, it is synchronously input into the native classification branch, localization branch, confidence branch of YOLOv8n, as well as the newly added lightweight target velocity prediction branch. The target velocity prediction branch shares the bottom-level features with the native branches and synchronously receives the pixel-physical space coordinate transformation matrix and the real-time operation parameters of the conveyor belt to predict the motion trajectory of the material target, the position and time of arrival at the sorting execution mechanism, and compensate for the model inference delay and the execution mechanism response delay. Then, the target detection results output by the native branches and the trajectory parameters output by the velocity prediction branch are synchronously input into the lightweight target matching unit. Through multi-dimensional inter-frame matching, a unique tracking identifier is assigned to each material target to achieve continuous target trajectory matching, avoid target misalignment and repeated detection under high-speed conveying, and finally complete the end-to-end joint output of detection and tracking.

[0038] In some examples, the improved YOLOv8n lightweight target detection model compares the action execution data fed back from the sorting execution stage with the detection results and sorting prediction results output by the model to identify and collect false detection and missed detection samples. After completing the automatic labeling of samples, incremental training is performed, and the model weights are updated online during the inference interval.

[0039] The above are merely preferred embodiments of this application. The scope of protection of this application is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in this application should be included within the scope of protection recorded in the claims.

Claims

1. A method for intelligent visual recognition and sorting control of large mining materials, characterized in that, The method includes: S1: Acquire the original image of the large ore material, and simultaneously collect the supplementary parameters corresponding to the original image. The supplementary parameters include: imaging exposure supplementary lighting parameters, field of view environment parameters, pixel physical space coordinate transformation matrix, and real-time operation parameters of the conveyor belt. S2: Input the original image and the supplementary parameters into the improved YOLOv8n model and output sorting control instructions. The improvement measures include: setting an adaptive pre-feature enhancement branch between the image input layer and the backbone network, setting a lightweight blind dehazing and deblurring fusion unit in the neck network, and integrating a spatiotemporal linkage detection and tracking head for velocity prediction and target matching. S3: Perform adaptive sorting operation based on the sorting control command.

2. The method according to claim 1, characterized in that, The pre-feature enhancement branch receives the imaging exposure supplementary lighting parameters, divides the input original image into regions, identifies and filters specular reflection areas without effective texture, performs adaptive dynamic range expansion processing on the remaining effective regions, performs directional texture feature enhancement on the dark areas, and outputs the fused feature map to the backbone feature extraction network.

3. The method according to claim 2, characterized in that, The pre-feature enhancement branch includes a reflection suppression submodule and a dark area enhancement submodule; The reflection suppression submodule is used to identify specular reflection areas, generate an invalid feature marker matrix, and suppress the feature weights of specular reflection areas. The dark area enhancement submodule extracts the lateral and longitudinal gradient features of the dark area of ​​large ore materials. It strengthens the strong gradient features corresponding to ore joints and suppresses the weak gradient features corresponding to homogeneous ore areas through learnable convolution weighting. The weighted gradient features and the grayscale features after dynamic range expansion are then concatenated and output.

4. The method according to claim 1, characterized in that, The blind dehazing and deblurring fusion unit receives the field of view environment parameters and performs defect repair on the multi-scale feature map output by the backbone feature extraction network.

5. The method according to claim 4, characterized in that, The blind dehazing and deblurring fusion unit includes a local fog concentration estimation subunit, a local motion blur degree estimation subunit, an adaptive weight allocation subunit, a dehazing branch, a deblurring branch, and a bypass direct access branch; The local fog concentration estimation subunit and the local motion blur estimation subunit are set in parallel, and their outputs are connected to the adaptive weight allocation subunit. The outputs of the adaptive weight allocation subunit are connected to the defogging branch, the deblurring branch, and the bypass direct branch, respectively. The adaptive weight allocation subunit allocates computational weights to the defogging branch and the deblurring branch of the corresponding region according to the fog concentration and motion blur estimation results of each local region of the feature map. For regions with significant defects, the corresponding branch is used to complete feature repair, and for regions without defects, the output is directly output through the bypass direct branch.

6. The method according to claim 1, characterized in that, The spatiotemporal linkage detection and tracking head adds a lightweight target velocity prediction branch and a lightweight target matching unit to the original classification branch, localization branch, and confidence branch of YOLOv8n; the velocity prediction branch shares the underlying features with the other branches, and its output is connected to the target matching unit.

7. The method according to claim 6, characterized in that, The target speed prediction branch receives the pixel physical space coordinate transformation matrix and the real-time operating parameters of the conveyor belt to predict the motion trajectory of the material target, the position and time of arrival at the sorting execution mechanism, and compensate for the model inference delay and the execution mechanism response delay; the lightweight target matching unit assigns a unique tracking identifier to each material target to achieve continuous trajectory matching between frames and avoid target misalignment and repeated detection under high-speed conveying.

8. The method according to claim 1, characterized in that, The improved YOLOv8n lightweight target detection model compares the action execution data fed back from the sorting execution stage with the detection results and sorting prediction results output by the model to identify and collect false detection and missed detection samples. After completing the automatic labeling of samples, incremental training is performed, and the model weights are updated online during the inference interval.