Deep learning-based network camera image sharpness enhancement method and system

Abstract

The present application relates to the technical field of underground network image definition enhancement, and particularly relates to a network camera image definition enhancement method and system based on deep learning. The method comprises the following steps: acquiring a network camera image of a mine underground operation area, and performing quality label classification processing on the image to construct a training sample containing multiple image features; analyzing the difference features of the training sample to determine the degradation features of the underground image; based on the degradation features, analyzing the structural loss damage degree of the key area of the repaired underground image to generate the structural definition of the image, training and parameter optimization of the deep learning model based on the structural definition, determining the image enhancement parameter configuration, and embedding it into the image processing module of the mine network camera system to realize real-time enhancement processing of the collected image; the present application enhances the network camera image definition to realize more efficient operation of the underground equipment.

Description

Technical Field

[0001] This invention relates to the field of network image sharpness enhancement technology in mining, and in particular to a method and system for enhancing network camera image sharpness based on deep learning. Background Technology

[0002] Network camera systems are playing an increasingly important role in mine safety monitoring, operational environment supervision, and intelligent inspection. However, the complex lighting conditions, dust, fog, and vibration interference in mines often result in insufficient clarity, low contrast, and blurred details in the acquired network camera images, severely impacting image quality and the accuracy of subsequent intelligent analysis. Traditional image enhancement methods rely heavily on techniques such as fixed filtering and histogram equalization. While these methods improve image contrast, they often come with side effects such as noise amplification, loss of detail, and structural distortion, making it difficult to meet the real-time, stable, and high-quality enhancement requirements for mine images. In recent years, with the rapid development of deep learning technology, image enhancement methods based on convolutional neural networks have gradually become a research hotspot in the field of image processing due to their powerful feature extraction capabilities and end-to-end learning advantages. Especially in low-light and complex environments, deep learning methods can adaptively recover the structural details and texture features of images, effectively improving image clarity and visual quality. However, current deep learning enhancement methods for mine network camera images still face many challenges. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for enhancing the image sharpness of network cameras based on deep learning, in order to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a deep learning-based network-based method for enhancing the sharpness of camera images includes the following steps:

[0005] Step S1: Obtain network camera images of the underground working area of ​​the mine, and perform quality label classification processing on the network camera images to construct training samples containing multiple image features;

[0006] Step S2: Analyze the difference features of the training samples to determine the degradation features of the training sample images under mining conditions;

[0007] Step S3: Analyze the degree of structural loss and damage in key areas of the underground image based on the degradation characteristics of the underground image, repair the degree of structural loss and damage in key areas, and generate the structural clarity of the underground image;

[0008] Step S4: Based on the structural clarity of the underground images, train and optimize the parameters to determine the image enhancement parameter configuration, and embed the structural clarity of the underground images into the image processing module deployed in the mine network camera system to perform real-time enhancement processing on the captured images and output a structurally clear image of the mine camera.

[0009] The present invention also provides a deep learning-based network camera image sharpness enhancement system for performing the deep learning-based network camera image sharpness enhancement method described above. The deep learning-based network camera image sharpness enhancement system includes:

[0010] The training sample acquisition module is used to acquire network camera images of underground mining areas, perform quality label classification on the network camera images, and construct training samples containing multiple image features.

[0011] The image degradation determination module is used to analyze the differential features of training samples and determine the degradation features of the training sample images.

[0012] The image restoration module is used to analyze the degree of structural loss and damage in key areas of underground images based on the degradation characteristics of underground images, and to restore the degree of structural loss and damage in key areas, thereby generating structural clarity of underground images.

[0013] The image enhancement processing module is used to train and optimize parameters based on the structural clarity of the underground images to determine the image enhancement parameter configuration. It also embeds the structural clarity of the underground images into the image processing module deployed in the mine network camera system, performs real-time enhancement processing on the camera-acquired images, and outputs a clear image of the mine camera structure.

[0014] (1) By collecting and classifying network camera images of the underground mining area through step S1, a training sample containing multiple image features is constructed, which fully covers the diversity and degradation features of underground mining images, providing a rich data foundation for subsequent differential feature analysis and effectively improving the representativeness and accuracy of the training sample.

[0015] (2) Step S2 extracts differential features based on training samples, accurately determines the degradation features of the underground image, and can distinguish different types of image degradation phenomena, such as blurring, noise, structural damage, etc., which helps to finely characterize the quality status of the underground image and improves the recognition accuracy and reliability of degradation features.

[0016] (3) By analyzing and repairing the degree of structural loss and damage in key areas through step S3, the structural clarity of the underground image is generated, which realizes the precise repair of complex structural degradation in the mining operation environment, enhances the integrity and detail of the image structure, provides an effective structural basis for subsequent enhancement processing, and significantly improves the image quality.

[0017] (4) Step S4 trains and optimizes the deep learning model based on structural clarity, determines the efficient image enhancement parameter configuration, and embeds it into the image processing module of the mine network camera system, realizing real-time enhancement processing of the acquired images, significantly improving the clarity and detail of the mine camera images, meeting the stringent requirements of the underground environment for real-time performance and stability, and improving the accuracy and reliability of mine safety monitoring and intelligent analysis. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a deep learning-based network-based method for enhancing the sharpness of camera images.

[0019] Figure 2 A schematic diagram of a deep learning-based network-based camera image sharpness enhancement system.

[0020] Figure 3 A schematic diagram of network camera images for underground mining equipment;

[0021] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0024] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0025] To achieve the above objectives, please refer to Figures 1 to 3 A deep learning-based network-based method for enhancing the sharpness of camera images includes the following steps:

[0026] Preferably, step S1: acquire network camera images of the underground working area of ​​the mine, perform quality label classification processing on the network camera images, and construct training samples containing multiple types of image features;

[0027] Optionally, step S1 includes the following steps:

[0028] Step S11: Acquire network camera images of the underground mining area, including original image frames and their timestamps;

[0029] Step S12: Denoise the network camera image;

[0030] Step S13: Identify key target regions based on network camera images;

[0031] Step S14: Label the key target area image with image quality tags for underground mining operations;

[0032] Step S15: Construct training samples containing multiple image features from the image quality labels.

[0033] In this embodiment, industrial-grade network cameras deployed at the main transport channels and intersecting roadways of the mine acquire raw image frame data of the underground operation scene. The cameras are equipped with infrared night vision and wide dynamic range capabilities, with an image resolution of 1920×1080, a sampling frequency of 1 frame per second, and automatically appended with UTC standard format timestamps via the image acquisition system. To address the common high ISO noise and dust interference in underground images, an improved nonlocal means (NLM) algorithm is used for image denoising, combined with a local brightness analysis strategy, to improve the overall signal-to-noise ratio while maintaining image edge clarity, thereby enhancing the robustness of subsequent recognition. Based on the YOLOv5s network model, target detection is performed on the denoised image frames to extract key operational areas, including miners, transport vehicles, and roadway support structures. Only areas with a confidence level higher than 0.6 are retained for subsequent training. The extracted target areas are quality-labeled using an image annotation platform. Image labels include five categories: "clear," "blurred," "occluded," "abnormal brightness," and "image noise." Manual annotation is combined with automatic feature extraction results for verification, ensuring a label accuracy rate higher than 95%. The grayscale distribution features, edge gradient maps, and texture features (such as the GLCM matrix) of each image are extracted as image features. These features are combined with quality labels to construct a training sample set, which serves as the training data source for the subsequent image sharpness recognition model.

[0034] In another embodiment, a 5G low-latency communication module is used to transmit image frames captured by the underground camera back to the ground data center in real time. The image resolution is 1280×720, and the sampling frequency is increased to 2 frames / second. Simultaneously, an RTC clock embedded within the camera provides timestamp information, facilitating accurate reconstruction of image sequences in scenarios with communication delays. During image preprocessing, the Retinex algorithm is introduced for brightness equalization, while bilateral filtering and adaptive median filtering are integrated to improve image clarity and contrast in low-light environments, making it more suitable for high-dust, low-illuminance scenarios such as coal mines. A lightweight MobileNet-SSD model is used for target area recognition, balancing recognition accuracy with edge computing resource consumption. It focuses on identifying specific structures such as underground signal lights, track bends, and ventilation facilities, enhancing the targeted analysis capability for communication anomalies. The number of image quality labels has been expanded to seven categories, with the addition of two special labels: "moisture interference" and "signal loss image," to address image distortion in high-humidity mines or communication interruption scenarios. A semi-supervised labeling mechanism is introduced into the labeling process, using clustering methods to assist manual verification and improve efficiency. The feature extraction method in the training samples introduces the ResNet base network to generate high-dimensional feature vectors. Combined with image brightness histogram, Fourier transform features and deep semantic embedding information, a composite sample integrating visual and temporal data is constructed for subsequent multi-task training, including image classification, anomaly recognition and image enhancement.

[0035] Preferably, step S14 includes the following steps:

[0036] Step S141: Enhance the local contrast of the key target region image and generate a target region structure map that reflects the sharpness of the image;

[0037] Step S142: Statistically analyze the edge response amplitude distribution of gradients in different directions of the target region structure map, and extract the gradient responses in the horizontal and vertical directions respectively;

[0038] Step S143: Calculate the average edge strength in the horizontal and vertical directions to form the average edge strength;

[0039] Step S144: Convert the target region structure map into an original grayscale image;

[0040] Step S145: Perform grayscale histogram statistical analysis on the original grayscale image, extract the grayscale fluctuation intensity, and generate a brightness quality feature vector;

[0041] Step S146: Divide the quality level labels of the key target region image based on the average edge intensity and brightness quality feature vector, wherein the labels include two levels: normal and abnormal;

[0042] Step S147: Use quality grade labels to annotate the key target area images with image quality labels for underground mining operations.

[0043] In this embodiment, the Adaptive Contrast Enhancement Algorithm (CLAHE) is used to locally process the extracted key target region image. By setting a contrast limit threshold and window size, the details, textures, and boundary contours in the image are enhanced to obtain a target region structure map. Subsequently, the Sobel operator is applied to extract the horizontal (Gx) and vertical (Gy) gradient responses of the structure map, respectively, to obtain the edge response amplitude distribution map, and the distribution of gradients in each direction in the image is statistically analyzed. Pixel-level mean calculations are performed on the gradient responses in the horizontal and vertical directions to obtain the average edge intensity in the horizontal and vertical directions, and their weighted average is calculated as the overall average edge intensity of the image. Then, the structure map is converted into a standard grayscale image using RGB channel weighted fusion for subsequent brightness characteristic analysis. A 256-level histogram analysis is performed on the grayscale image to extract statistical features such as variance, skewness, and entropy of the grayscale values, forming a brightness quality feature vector. The average edge intensity and the brightness feature vector are input into a trained binary classification model to output the quality level label of the key target region image, divided into "normal" and "abnormal". Finally, the image quality label is completed based on the classification results.

[0044] In another embodiment, a Retinex-based illumination estimation enhancement algorithm is used to enhance local image contrast, emphasizing the restoration of texture details in dark areas and adapting to low-light scenes such as coal mines. The output target region structure map maintains brightness uniformity while improving image boundary recognition. During gradient extraction, in addition to the conventional Sobel operator, the Prewitt operator is used in parallel to increase robustness, and a directional weight merging mechanism is employed to generate a directional response map, reducing the impact of underground dust and fog interference on edge recognition. The mean and standard deviation of the gradient responses in the horizontal and vertical directions are calculated to form a multidimensional edge strength index, which not only quantifies image sharpness but also reflects the level of edge consistency. During image conversion, the Y component of the brightness channel (YCrCb color space) is preserved to avoid loss of brightness features under non-uniform illumination conditions, making subsequent grayscale statistics more representative. Brightness entropy is introduced as an additional feature dimension, and combined with the low-order and high-order moments of the image grayscale distribution, a 5-dimensional brightness quality feature vector is constructed to adapt to different types of image quality degradation modes. Label classification is performed using threshold rules trained based on a decision tree algorithm. A joint judgment logic of grayscale fluctuation and edge contrast is introduced into the model to improve the recognition rate of abnormal images, ensuring that identified "abnormal" images receive higher processing priority. Label assignment is completed using a combination of automatic annotation and manual verification. A manual review interface is provided for edge-judgment images to improve label quality, reduce the risk of misjudgment, and provide accurate supervision data for subsequent deep learning models.

[0045] Preferably, step S2: analyze the difference features of the training samples and determine the degradation features of the training sample images in the mining process;

[0046] Optionally, step S2 includes the following steps:

[0047] Step S21: Extract abnormal image degradation response features from target region images with abnormal image quality labels in the training samples;

[0048] Step S22: Extract image structural integrity features from target region images with normal image quality labels in the training samples;

[0049] Step S23: Analyze the differences between image abnormal degradation response features and image structural integrity features;

[0050] Step S24: Determine the degradation features of the mining images of the training sample images based on the difference features between the training images.

[0051] In this embodiment, key target region images labeled "abnormal" are selected as the abnormal sample set, and spatial and frequency domain feature analysis is performed on the images in this set. Response features of the images under abnormal degradation are extracted through multiple dimensions, including Gaussian blur response, local contrast distribution, image gradient direction consistency, and edge fracturing. These features can reflect problems such as decreased image sharpness, structural damage, or illumination imbalance. Structural integrity analysis is performed on target region images labeled "normal," extracting strong structural features such as image edge continuity, texture stability, gray-level uniformity, and brightness contrast stability. A structural integrity feature space reflecting the normal image state is constructed by statistically analyzing the image's gradient direction histogram distribution, edge density, and texture entropy. A multi-dimensional feature comparison analysis method is used to compare the two feature sets obtained in steps S21 and S22 dimension by dimension, calculating the difference index of each feature dimension between "normal" and "abnormal" images. The difference index includes statistics such as Euclidean distance, KL divergence, and mutual information difference, used to measure the significant changes in the corresponding features of the image before and after degradation. Significantly different features are used as candidate degradation markers. The difference feature vectors obtained in step S23 are input into the feature selection model. Using a combination of information gain and principal component analysis (PCA), the key degradation feature dimensions that best represent the degradation of the underground image quality are selected to establish an underground image degradation feature template, which is then used to construct the input features for the subsequent image quality prediction model.

[0052] Preferably, step S21 includes the following steps:

[0053] Step S211: For target region images with abnormal image quality labels in the training samples, identify regions with abnormal brightness distribution in the target region images and perform brightness normalization processing on the target region images.

[0054] Step S212: Statistically analyze the distribution of structural fuzziness in the image;

[0055] Step S213: For the case of structural fuzziness, extract the gradient direction information of the image and count the distribution ratio of gradients in different directions;

[0056] Step S214: Calculate the directional distribution discreteness in the local area using the distribution ratio of gradients in different directions. When it is greater than the preset perturbation judgment threshold, it is used as the image abnormal degradation response feature.

[0057] In one embodiment, target region images labeled "abnormal" in the training samples are preprocessed to identify regions with abnormal brightness distribution. A combination of local contrast analysis and grayscale histogram equalization is used to identify overexposed, underexposed, or unevenly lit image sub-regions. Brightness normalization is then applied to the entire image to suppress the impact of brightness interference on structural feature extraction. Subsequently, the structural blur distribution in the normalized image is statistically analyzed. Specifically, the blur degree of each region in the image is evaluated by jointly using a gradient magnitude map and an edge sharpness mapping map to construct a structural blur heatmap. For structurally blurred regions, gradient direction information is extracted from the local region. The Sobel operator is used to calculate the gradient components in the horizontal and vertical directions, and the pixel ratio of gradients in each direction is statistically analyzed to construct a direction histogram. Based on the direction histogram, the dispersion of gradient direction distribution in the local region is calculated. Information entropy and range indices are used to measure the dispersion. When the dispersion of the local direction distribution exceeds a preset perturbation judgment threshold (e.g., 0.85), the feature corresponding to that region is marked as an image abnormal degradation response feature for subsequent difference modeling.

[0058] In another embodiment, a multi-scale Retinex algorithm is used for brightness normalization. Local brightness mean values ​​are extracted from receptive fields of 3×3, 5×5, and 7×7 respectively, and the results are fused to achieve smoother brightness adjustment and enhance adaptability to areas of abnormal brightness. Structural fuzziness distribution is calculated using a fuzzy response map based on the Laplacian operator. The image is segmented using a variance threshold to identify fuzzy regions and generate corresponding fuzzy masks. Based on this, the Scharr operator is used instead of traditional gradient extraction methods to obtain more refined gradient direction information in the image. The pixel ratio for each direction (0°, 45°, 90°, 135°) is statistically analyzed to construct a direction distribution feature vector. The standard deviation and normalized entropy of the local direction distribution are calculated to determine its directional dispersion. If the deviation exceeds a dynamically set perturbation threshold (adaptively adjusted based on the overall image noise level), the directional features of that local region are identified as abnormal degradation response features and recorded.

[0059] Preferably, step S22 includes the following steps:

[0060] Step S221: Perform illumination equalization processing on the target region images with normal image quality labels in the training samples to preserve the original structural details of the images;

[0061] Step S222: Detect the direction of continuous edges in the target region image, calculate the linear extension of the edge pixel distribution, and form an edge continuity feature map;

[0062] Step S223: Extract the gray-level gradient distribution of the target region image, calculate the difference between the gradient change range and the mean within the region, and generate image detail intensity features;

[0063] Step S224: Extract the geometric outer contour of the device target based on the image detail intensity features, and determine the device contour structure parameters;

[0064] Step S225: Use the edge continuity feature map, image detail intensity feature and device contour structure parameters as image structural integrity features.

[0065] In one embodiment, illumination equalization processing is performed on target area images with the image quality label "normal" in the training samples. A combination of multi-channel histogram matching and local brightness compensation is used to balance the brightness distribution of different regions of the image, while preserving the original edge details and structural texture information of the image during processing. Next, edge detection is performed on the illumination-equalized image. The Canny edge detection algorithm is used to identify edge lines with coherent structures in the image, and an edge continuity feature map is constructed based on edge direction consistency and edge pixel density. A linear fitting algorithm is used to calculate the extension degree of edge pixels in the main direction, thereby measuring the coherence of the image structure. Further, grayscale gradient analysis is performed on the image. The Sobel operator is used to extract grayscale gradient maps in the horizontal and vertical directions of the image, and the gradient change range and gradient mean difference are statistically analyzed in multiple sub-regions to form a detail intensity feature map of the image, reflecting the degree of local structural changes in the image. Based on this, combined with the image detail intensity features, the outer contour of the mining equipment in the image is extracted. Geometric fitting is performed using parameters such as edge closure, contour symmetry, and area ratio to obtain the contour structure parameters of the equipment target. Finally, the edge continuity feature map, image detail intensity feature, and device contour structure parameters obtained above are combined as the structural integrity feature of the image, which is used for comparative analysis with abnormal image features.

[0066] In another embodiment, image illumination equalization is enhanced using the MSRCR algorithm based on Retinex theory. This algorithm processes the original image through multi-scale dynamic range compression and color restoration functions, improving local contrast while preserving structural edges. Subsequently, a combination of edge linking algorithms and Hough transform is used to detect the direction of continuous edges in the image. After identifying the main edge directions, their projection density on the image's principal axis is calculated as an edge extension index, constructing an edge continuity map. For image detail feature extraction, a Laplacian-enhanced grayscale gradient map is used. By calculating the difference between the maximum and minimum gradient values ​​within a sliding window and combining this with local mean differences, a feature map reflecting the intensity of image texture details is constructed. Device contour extraction employs a morphological gradient and contour tracking algorithm. Continuous edges are closed, and after extracting the closed contour region, its area, perimeter, roundness, and principal axis tilt angle are calculated. Edge continuity, image detail intensity, and device contour structural parameters are fused to form an image structural integrity feature vector for the normal image.

[0067] Preferably, step S23 includes the following steps:

[0068] Step S231: Perform spatial registration on the image abnormal degradation response features and image structural integrity features, and filter out comparable image corresponding region data;

[0069] Step S232: Extract the spatial coordinate sequence of edge pixels for the corresponding region of the image, and calculate the gradient magnitude change rate between adjacent edge pixels.

[0070] Step S233: Calculate the edge structure consistency offset parameter and generate the edge structure deviation map;

[0071] Step S233: Analyze the texture perturbation features of the edge structure deviation map;

[0072] Step S234: Using the corresponding region data of the image, extract the contour closure feature of the geometric structure of the underground equipment and calculate the contour morphology inconsistency parameter;

[0073] Step S235: Fuse the difference index including texture perturbation features and contour shape inconsistency parameters, determine whether the difference index exceeds the preset difference threshold, and generate difference features.

[0074] In one embodiment, the extracted image anomaly degradation response features and image structural integrity features are spatially registered. Comparable image regions are selected through scale matching and image region mapping. Based on these regions, a two-dimensional spatial coordinate sequence of edge pixels is extracted, and the gray-level gradient amplitude change rate between adjacent edge pixels is further statistically analyzed to construct an edge gradient change sequence. An edge structure consistency offset parameter is calculated based on this sequence, using edge orientation similarity and local change rate as the calculation basis to construct an edge structure deviation map. Higher values ​​in the map indicate greater differences in edge structure at the corresponding location. Based on this, texture perturbation analysis is performed on the edge structure deviation map. Combining the frequency of local texture direction changes and the density of detail mutations, texture perturbation features caused by structural anomalies in the image are extracted. Using spatially registered image region data, the geometric structure contours of underground mining equipment within the corresponding region are extracted. Contour closure features are constructed using contour closure degree and boundary smoothness. The degree of difference between each feature is calculated to obtain contour morphology inconsistency parameters. Texture perturbation features and contour morphology inconsistency parameters are fused to construct an image structure difference index. It is then determined whether the index exceeds a preset difference threshold. If it does, difference features are generated and used for subsequent image degradation model training.

[0075] In another embodiment, the corresponding regions of the image anomaly and normal samples are registered using an affine transformation based on SURF feature points, and image pairs with structural alignment higher than a set threshold (e.g., 0.92) are selected using a region consistency metric function. Subsequently, edge pixel coordinates are extracted from the image pairs, differential filtering is used to enhance their gradient response, and the gradient magnitude change rate of adjacent edge pixels is statistically analyzed to extract local edge perturbation sequences. By statistically analyzing the average fluctuation amplitude and maximum slope of the edge perturbation sequences, edge structure consistency offset parameters are constructed, and an edge structure deviation map is generated. Wavelet texture decomposition is used to perform multi-scale texture analysis on the deviation map, extracting the local texture perturbation energy spectrum as the image texture perturbation feature. Simultaneously, combining the symmetry and connectivity features of the device's outer contour boundary, contour morphology inconsistency parameters are calculated to characterize the degree of geometric mismatch in the anomaly image. The texture perturbation features and contour morphology inconsistency parameters are then fused after unified normalization. A comprehensive difference evaluation function is used to determine whether the difference index exceeds a system-set threshold; if it does, it is determined to be a valid difference feature.

[0076] Preferably, step S24 includes the following steps:

[0077] Step S241: Statistically analyze the degradation response of the differential characteristics;

[0078] Step S242: Analyze the performance of image degradation based on the degradation response;

[0079] Step S243: Extract the orientation change of the structure in the image using the difference features, and calculate the angle offset between the main edge direction and the reference edge direction in the target area as the structure response deflection angle;

[0080] Step S244: Calculate the local contrast reduction value of the structural response deflection angle;

[0081] Step S245: Combine the local contrast reduction value and the image degradation performance to determine the image degradation characteristics of the training sample images.

[0082] In one embodiment, the differential features extracted in step S23 are statistically analyzed. Based on the intensity, density, and distribution range of different differential features, the corresponding degradation response is determined, including increased image blur, broken structural edges, and enhanced texture perturbation, thus constructing a degradation response index sequence. Subsequently, based on this degradation response index sequence, the overall image degradation is classified and analyzed, distinguishing it into different types such as edge degradation-dominated, brightness perturbation-dominated, and composite degradation, as manifestations of image degradation. Furthermore, the differential features extracted from the image are used to analyze the edge direction of key structural regions, extracting the main edge direction and the reference edge direction of the corresponding region in the reference image, and calculating the angle offset between them. This angle is the structural response deflection angle, used to characterize the degree of structural distortion. Then, the local contrast changes around the structural response deflection angle region are statistically analyzed, and the contrast decrease value is calculated using a window sliding method to characterize the degree of detail loss in this region. The local contrast decrease value of the structural response deflection angle is fused with the overall performance of image degradation and a weighted normalization model is used to form a unified image degradation characterization feature, which is used as the mining image degradation feature of the training sample image for model construction and subsequent recognition and judgment.

[0083] In another embodiment, the differential features are classified and statistically analyzed using multidimensional clustering. The aggregation degree of texture perturbation features, edge deviation features, and contour inconsistency features is statistically analyzed to form a degradation response intensity distribution map, which is used to evaluate the image degradation performance. Combining the regional clustering characteristics of degradation performance, the main degradation types of the image are determined, and corresponding degradation performance codes are constructed. In image structural response analysis, the Canny edge detection algorithm is used to finely extract the main edges. The Hough transform is used to fit the main edge direction line segments, and the angle between these segments and the standard edge direction in the reference image is matched to obtain the structural response deflection angle. Subsequently, the local contrast Laplacian algorithm is applied within the deflection region to calculate the contrast decrease value, which is used to evaluate the actual impact of degradation in this region on visible details. The degradation performance code, structural response deflection angle, and local contrast decrease value are uniformly input into a multidimensional fusion function, and principal component analysis is used to extract its comprehensive features, obtaining the mining image degradation features of the training sample images.

[0084] Preferably, step S3: Analyze the degree of structural loss and damage in key areas of the underground image based on the degradation characteristics of the underground image, repair the degree of structural loss and damage in key areas, and generate the structural clarity of the underground image;

[0085] Of particular importance, step S3 includes the following steps:

[0086] Step S31: Identify the structural loss target region in the degradation features of the underground image;

[0087] Step S32: Identify the structural damage and destruction status of key areas within the target area of ​​structural loss;

[0088] Step S33: Repair the structural damage in critical areas;

[0089] Step S34: Analyze whether the repaired area meets the preset restoration standard. If it does, the output is the structural clarity of the underground image.

[0090] In one embodiment, step S3 includes the following processing: Based on the degradation features of the mine image determined in step S2, the system performs regional structural evaluation on the image to be processed, identifying structurally blurred and detail-lacking target regions in the image. Subsequently, within the identified structurally damaged target regions, the system further analyzes the structural features of key regions in the image, including edge integrity, contour connectivity, and texture continuity, to identify the structural damage status of the key regions. For the identified structurally damaged areas, an image inpainting model based on image context structure reconstruction is used for reconstruction processing, combined with a deep learning feature generator to complete the detail information of the key regions, enhancing the structural consistency of the regions. After the repair is completed, the system performs a structural consistency check on the repaired region and calculates the structural matching degree between the restored region and the neighboring images. If the restoration effect meets the preset restoration criteria, the repair is deemed effective, and the structural clarity index corresponding to the image is output. This structural clarity serves as a basic reference feature for subsequent image enhancement network training and deployment.

[0091] In another embodiment, step S3 accurately locates the structural loss target region using a residual prediction module that integrates multi-scale features. A guided attention mechanism is employed to enhance the model's focus on key structural regions. After identifying the damage to key structures, an edge-preserving autoencoder-based insulation network is introduced for structural reconstruction. Simultaneously, multi-dimensional texture alignment is performed on the image to ensure consistency between the repaired region and the original image's structural style. The system further utilizes a structural contrast loss function to quantitatively evaluate the restoration quality of the repaired region. When the detection result meets both structural fidelity and sharpness thresholds, the image is considered structurally sharp, and a structural sharpness index is output for image enhancement processing.

[0092] Of particular importance, step S33 includes the following steps:

[0093] Step S331: Calculate the lateral and longitudinal edge lengths of the structural damage in the critical area;

[0094] Step S332: Calculate the horizontal and vertical image ratio deviation using the horizontal edge length and the vertical edge length;

[0095] Step S333: Adjust the pixel arrangement of the key repair area to correct the horizontal and vertical image ratio deviation;

[0096] Step S334: Remove noise indicating structural damage or destruction in critical areas.

[0097] In one embodiment, the system extracts edge features from the identified key areas of structural damage, statistically analyzing the horizontal and vertical edge lengths of each area. The spatial distribution of these edges identifies the directional characteristics of the structural damage. Based on the edge lengths, the system further calculates the ratio between the horizontal and vertical edge lengths, forming a horizontal-vertical image ratio deviation index. Next, the system adjusts the pixel arrangement of the key repair areas according to this ratio deviation, ensuring directional consistency and regularity in the image during structural repair, thus enhancing the restoration effect. To further improve the quality of the repaired image, the system also removes noise, shadows, and other impurities in the repair area. It employs image smoothing techniques based on local texture statistics to eliminate discrete, isolated, unstructured pixels, effectively removing interference information during structural repair. These steps improve the accuracy of structural repair and the clarity of the regional image, providing structural integrity assurance for subsequent image enhancement processing.

[0098] In another embodiment, the system utilizes an edge detection operator to extract directional contours from images of structurally damaged key regions. It separates the main structural edges using an edge response intensity threshold and measures their horizontal and vertical edge lengths respectively. Combining the edge length differences, an image scaling deviation model is constructed. This deviation is used as a constraint for pixel restoration, and the pixel arrangement in the lost region is reconstructed in a regularized manner to ensure that the restored image maintains the spatial structure of the original image. Figure 1 Consistency was then established. Subsequently, image median filtering and high-frequency interference detection algorithms were used to remove noise from the repaired area, eliminating isolated high-frequency abnormal pixels and non-connected texture fragments, thereby improving the overall texture continuity and edge clarity of the image, enhancing the recognizability of the image structure, and improving the stability of subsequent processing.

[0099] Preferably, step S4: Based on the structural clarity of the underground images, training and parameter tuning are performed to determine the image enhancement parameter configuration, and the structural clarity of the underground images is embedded into the image processing module deployed in the mine network camera system to perform real-time enhancement processing on the captured images and output a structurally clear image of the mine camera.

[0100] Optionally, step S4 includes the following steps:

[0101] Step S41: Construct an image structure feature set based on the structural clarity of the underground images;

[0102] Step S42: Use the image structure feature set as input to the deep learning enhancement network, perform fusion optimization training, and determine the image enhancement parameter configuration;

[0103] Step S43: The obtained image enhancement parameters are configured and embedded into the image processing module deployed in the mine network camera system. Combined with the structural response information of the images acquired by the camera system, the real-time enhancement processing of the image structural details is completed.

[0104] Step S44: Perform consistency detection and degradation residual correction on the enhanced image structural features to output a clear image of the mine camera structure.

[0105] In one embodiment, based on the underground image structure information extracted in the aforementioned steps, an image structure feature set is constructed. This feature set includes edge intensity distribution, contour continuity parameters, grayscale gradient change values, and texture direction consistency, comprehensively characterizing the structural clarity of the image. Based on this, the image structure feature set is used as input to construct an adapted deep learning enhancement network structure. This network includes multiple feature fusion layers, channel attention modules, and structural residual feedback units. Supervised learning is employed to perform fusion optimization training on the network to determine image enhancement parameter configurations suitable for underground images. Subsequently, the trained image enhancement parameter configuration is embedded into the image processing module deployed in the mine network camera system. Combined with the image structure response information generated in real-time during the camera system's acquisition process, real-time enhancement processing of the structural details of key target areas is achieved. The enhanced image structure features undergo consistency detection and degradation residual correction. The consistency detection includes frame-to-frame structural alignment and local detail residual analysis to ensure the structural coherence and detail integrity of the output image, resulting in a clear structural image from the mine camera.

[0106] In another embodiment, the construction of the image structural feature set further includes a low-contrast region detection and blurred region localization module to improve the local sensitivity of structural features. The deep learning enhancement network employs a U-Net structure combined with residual connections, and a structural loss function is introduced during training to enhance the model's specificity and robustness in image detail restoration. After being deployed to the image processing module, the image enhancement parameters can be dynamically adjusted based on the image acquisition frame rate and ambient lighting conditions to achieve low-latency adaptive enhancement processing. After enhancement processing, the feature residual maps from historical image frames are further combined to perform dynamic residual correction on the current image structural features, repairing existing edge jaggedness or over-enhanced details, thereby obtaining a clear image of the mine camera structure with strong structural consistency and complete detail preservation.

[0107] The present invention also provides a deep learning-based network camera image sharpness enhancement system for performing the deep learning-based network camera image sharpness enhancement method described above. The deep learning-based network camera image sharpness enhancement system includes:

[0108] The training sample acquisition module 101 is used to acquire network camera images of the underground working area of ​​the mine, and to perform quality label classification processing on the network camera images to construct training samples containing multiple image features.

[0109] The image degradation determination module 102 is used to analyze the difference features of the training samples and determine the degradation features of the training sample images.

[0110] The image restoration module 103 is used to analyze the degree of structural loss and damage in key areas of the underground image based on the degradation characteristics of the underground image, and to restore the degree of structural loss and damage in key areas to generate the structural clarity of the underground image.

[0111] The image enhancement processing module 104 is used to train and optimize parameters based on the structural clarity of the underground mine image to determine the image enhancement parameter configuration, and to embed the structural clarity of the underground mine image into the image processing module deployed in the mine network camera system. It performs real-time enhancement processing on the camera-acquired images and outputs a structurally clear image of the mine. The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for enhancing the sharpness of network camera images based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain network camera images of the underground working area of ​​the mine, and perform quality label classification processing on the network camera images to construct training samples containing multiple image features; Step S2: Analyze the difference features of the training samples to determine the degradation features of the training sample images; wherein, step S2 includes the following steps: Step S21: Extract abnormal image degradation response features from target region images with abnormal image quality labels in the training samples; Step S22: Extract image structural integrity features from target region images with normal image quality labels in the training samples. Step S22 includes the following steps: Step S221: Perform illumination equalization processing on the target region images with normal image quality labels in the training samples to preserve the original structural details of the images; Step S222: Detect the direction of continuous edges in the target region image, calculate the linear extension of the edge pixel distribution, and form an edge continuity feature map; including edge detection on the image after illumination equalization, identifying edge lines with continuous structure in the image, and constructing an edge continuity feature map based on edge direction consistency and edge pixel density; Step S223: Extract the gray-level gradient distribution of the target region image, calculate the difference between the gradient change range and the mean within the region, and generate image detail intensity features; Step S224: Extract the geometric outer contour of the device target based on the image detail intensity features, and determine the device contour structure parameters; Step S225: Use the edge continuity feature map, image detail intensity features, and device contour structure parameters as image structural integrity features; Step S23: Analyze the differences between image abnormal degradation response features and image structural integrity features; Step S24: Determine the degradation features of the training sample images based on the difference features between the training images, wherein step S24 includes the following steps: Step S241: Statistically analyze the degradation response of the differential characteristics; Step S242: Analyze the performance of image degradation based on the degradation response; Step S243: Extract the orientation change of the structure in the image using the difference features, and calculate the angle offset between the main edge direction and the reference edge direction in the target area as the structure response deflection angle; Step S244: Calculate the local contrast reduction value of the structural response deflection angle; Step S245: Combine the local contrast reduction value and the image degradation performance to determine the mining image degradation characteristics of the training sample images; Step S3: Analyze the degree of structural loss and damage in key areas of the underground image based on the degradation characteristics of the underground image, repair the degree of structural loss and damage in key areas, and generate the structural clarity of the underground image; Step S4: Based on the structural clarity of the underground images, train and optimize the parameters to determine the image enhancement parameter configuration, and embed the structural clarity of the underground images into the image processing module deployed in the mine network camera system to perform real-time enhancement processing on the captured images and output a structurally clear image of the mine camera.

2. The method for enhancing the sharpness of network camera images based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Acquire network camera images of the underground mining area, including original image frames and their timestamps; Step S12: Denoise the network camera image; Step S13: Identify key target regions based on network camera images; Step S14: Label the key target area image with image quality tags for underground mining operations; Step S15: Construct training samples containing multiple image features from the image quality labels.

3. The method for enhancing the sharpness of network camera images based on deep learning according to claim 2, characterized in that, Step S14 includes the following steps: Step S141: Enhance the local contrast of the key target region image and generate a target region structure map that reflects the sharpness of the image; Step S142: Statistically analyze the edge response amplitude distribution of gradients in different directions of the target region structure map, and extract the gradient responses in the horizontal and vertical directions respectively; Step S143: Calculate the average edge intensity in the horizontal and vertical directions to form the average edge intensity; which includes: calculating the pixel-level mean of the gradient response in the horizontal and vertical directions respectively to obtain the average edge intensity in the horizontal and vertical directions, and calculating their weighted average value as the overall average edge intensity of the image. Step S144: Convert the target region structure map into an original grayscale image; Step S145: Perform grayscale histogram statistical analysis on the original grayscale image, extract the grayscale fluctuation intensity, and generate a brightness quality feature vector; Step S146: Divide the quality level labels of the key target region image based on the average edge intensity and brightness quality feature vector, wherein the labels include two levels: normal and abnormal; Step S147: Use quality grade labels to annotate the key target area images with image quality labels for underground mining operations.

4. The method for enhancing the sharpness of network camera images based on deep learning according to claim 1, characterized in that, Step S21 includes the following steps: Step S211: For target region images with abnormal image quality labels in the training samples, identify regions with abnormal brightness distribution in the target region images and perform brightness normalization processing on the target region images. Step S212: Statistically analyze the structural fuzzy distribution in the image, including: statistically analyzing the structural fuzzy distribution in the normalized image; Step S213: For the case of structural fuzziness, extract the gradient direction information of the image and count the distribution ratio of gradients in different directions; Step S214: Calculate the directional distribution discreteness in the local area using the distribution ratio of gradients in different directions. When it is greater than the preset perturbation judgment threshold, it is used as the image abnormal degradation response feature.

5. The method for enhancing the sharpness of network camera images based on deep learning according to claim 1, characterized in that, Step S23 includes the following steps: Step S231: Perform spatial registration on the image abnormal degradation response features and image structural integrity features, and filter out comparable image corresponding region data; Step S232: Extract the spatial coordinate sequence of edge pixels for the corresponding region of the image, and calculate the gradient magnitude change rate between adjacent edge pixels; Step S233: Calculate the edge structure consistency offset parameter and generate an edge structure deviation map; including: statistically analyzing the grayscale gradient amplitude change rate between adjacent edge pixels, constructing an edge gradient change sequence, and calculating the edge structure consistency offset parameter based on the sequence. Step S233: Analyze the texture perturbation features of the edge structure deviation map; Step S234: Using the corresponding region data of the image, extract the contour closure feature of the geometric structure of the underground equipment and calculate the contour morphology inconsistency parameter; Step S235: Fuse the difference index including texture perturbation features and contour shape inconsistency parameters, determine whether the difference index exceeds the preset difference threshold, and generate difference features.

6. The method for enhancing the sharpness of network camera images based on deep learning according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Construct an image structure feature set based on the structural clarity of the underground images; Step S42: Use the image structure feature set as input to the deep learning enhancement network, perform fusion optimization training, and determine the image enhancement parameter configuration; Step S43: The obtained image enhancement parameters are configured and embedded into the image processing module deployed in the mine network camera system. Combined with the structural response information of the images acquired by the camera system, the real-time enhancement processing of the image structural details is completed. Step S44: Perform consistency detection and degradation residual correction on the enhanced image structural features to output a clear image of the mine camera structure.

7. A network camera image sharpness enhancement system based on deep learning, characterized in that, For performing the deep learning-based network camera image sharpness enhancement method as described in claim 1, the deep learning-based network camera image sharpness enhancement system includes: The training sample acquisition module is used to acquire network camera images of underground mining areas, perform quality label classification on the network camera images, and construct training samples containing multiple image features. The image degradation determination module is used to analyze the differential features of training samples and determine the degradation features of the training sample images. The image restoration module is used to analyze the degree of structural loss and damage in key areas of underground images based on the degradation characteristics of underground images, and to restore the degree of structural loss and damage in key areas, generating structural clarity images of underground images. The image enhancement processing module is used to train and optimize parameters based on the structural clarity of underground images to determine the image enhancement parameter configuration, and to embed the structural clarity of underground images into the image processing module deployed in the mine network camera system, to perform real-time enhancement processing on the camera-acquired images, and to output structurally clear images of the mine cameras.

Images