An image enhancement processing system based on deep learning

Abstract

The application belongs to the technical field of image processing, and discloses an image enhancement processing system based on deep learning, which comprises a degraded image acquisition module, an image degradation feature analysis module, a multi-scale deep enhancement module, a detail texture reconstruction module, a color brightness adaptive correction module and an enhanced result feedback optimization module. The image degradation feature analysis, multi-scale deep enhancement, residual learning, texture attention mechanism, color brightness joint correction and feedback optimization principles are used to identify the degradation type of the image to be enhanced, to perform layered enhancement on the brightness structure details, to reconstruct the detail texture and to correct the quality in a closed loop, so that the degraded area differentiation enhancement, dark area brightness recovery, main structure enhancement, edge detail reservation, color natural correction and stable output of the enhanced result are realized, and the problems of insufficient dark area, overexposure of bright area, noise amplification, texture distortion, color deviation and lack of feedback optimization caused by the unified processing of the traditional image enhancement mode are solved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to an image enhancement processing system based on deep learning. Background Technology

[0002] With the development of applications such as security monitoring, intelligent transportation, industrial inspection, medical imaging, mobile terminal shooting, and remote sensing image analysis, image data plays an increasingly important role in target recognition, defect detection, behavior analysis, and visual decision-making. However, in actual imaging processes, images to be processed are easily affected by factors such as low illumination, haze, motion blur, noise interference, overexposure, low contrast, and color shift, resulting in unclear target edges, loss of texture details, color distortion, and insufficient visibility in local areas, thus affecting subsequent visual recognition and analysis results. Therefore, a deep learning-based image enhancement processing system is needed to perform degradation feature analysis, multi-scale enhancement, detail texture reconstruction, color and brightness correction, and enhancement result feedback optimization on the image to be enhanced, in order to improve the stability and adaptability of image enhancement processing.

[0003] The following technical problems exist in the existing technology and need to be improved: (1) Existing image enhancement processing techniques mostly use fixed enhancement rules or single enhancement parameters to uniformly process the whole image. They lack joint analysis of degradation features such as brightness distribution, noise distribution, blur degree, contrast degradation and color shift, making it difficult to perform differentiated enhancement for different degradation areas. This can easily lead to problems such as insufficient enhancement of dark areas, overexposure of bright areas, amplification of noise and unnatural local contrast.

[0004] (2) Existing image enhancement processing techniques usually focus on brightness enhancement or contrast enhancement, lacking layered enhancement of low-frequency brightness information, mid-frequency structural information and high-frequency edge information. They also lack detailed texture reconstruction, color and brightness joint correction and enhancement result feedback optimization mechanisms, which can easily lead to edge breakage, texture distortion, color shift and over-enhancement in the enhanced image. They are difficult to adapt to the image enhancement needs under different shooting environments, different imaging device parameters and different degradation types. Summary of the Invention

[0005] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based image enhancement processing system. Utilizing image degradation feature analysis, multi-scale depth enhancement, residual learning, texture attention mechanisms, joint color and brightness correction, and feedback optimization principles, the system performs degradation type identification, brightness, structural detail layered enhancement, detail and texture reconstruction, and quality closed-loop correction on the image to be enhanced. This achieves differentiated enhancement of degraded areas, brightness restoration of dark areas, enhancement of the main structure, preservation of edge details, natural color correction, and stable output of enhancement results. It solves the problems of insufficient dark areas, overexposure of bright areas, noise amplification, texture distortion, color shift, and lack of feedback optimization caused by the uniform processing of traditional image enhancement methods.

[0006] This invention provides a deep learning-based image enhancement processing system, including a degraded image acquisition module, an image degradation feature analysis module, a multi-scale depth enhancement module, a detail texture reconstruction module, a color and brightness adaptive correction module, and an enhancement result feedback optimization module.

[0007] The degraded image acquisition module acquires the image to be enhanced and collects the shooting environment data, imaging device parameters and basic image attribute data corresponding to the image to be enhanced, forming degraded image input data; The image degradation feature analysis module receives degraded image input data and performs brightness distribution analysis, noise distribution analysis, blur degree analysis, contrast degradation analysis, and color shift analysis on the image to be enhanced, extracting the image degradation feature matrix. The multi-scale depth enhancement module receives the image degradation feature matrix and the image to be enhanced, constructs a multi-scale enhancement branch based on a deep convolutional neural network, performs low-frequency brightness restoration, mid-frequency structure enhancement and high-frequency edge enhancement on the image to be enhanced, and obtains low-frequency enhancement features, mid-frequency enhancement features and high-frequency enhancement features, and generates a multi-scale enhancement feature map based on the low-frequency enhancement features, mid-frequency enhancement features and high-frequency enhancement features. The detail texture reconstruction module receives multi-scale enhanced feature maps and, based on residual learning and texture attention mechanisms, reconstructs edge contours, fine textures, and weak contrast regions in the image to be enhanced, generating a detail texture reconstruction image. The color brightness adaptive correction module receives the detailed texture reconstruction image and performs brightness equalization, color shift correction and local contrast adjustment on the detailed texture reconstruction image according to the brightness distribution features and color shift features in the image degradation feature matrix to generate an enhanced output image. The enhancement result feedback optimization module receives the enhanced output image, performs noise residue assessment, detail fidelity assessment, color naturalness assessment, and over-enhancement risk assessment on the enhanced output image to obtain quality assessment results, and performs feedback corrections on the multi-scale depth enhancement module, detail texture reconstruction module, and color brightness adaptive correction module based on the quality assessment results.

[0008] Furthermore, the degraded image acquisition module includes image acquisition to be enhanced, shooting environment data acquisition, imaging device parameter acquisition, image basic attribute data acquisition, and data output.

[0009] Furthermore, the image degradation feature analysis module includes brightness distribution analysis, noise distribution analysis, blur degree analysis, contrast degradation analysis, color shift analysis, and image degradation feature matrix construction.

[0010] Furthermore, the multi-scale depth enhancement module includes multi-scale input construction, low-frequency brightness restoration, mid-frequency structure enhancement, high-frequency edge enhancement, multi-scale enhancement feature fusion, degradation-guided weight calculation, and multi-scale enhancement feature map output.

[0011] Furthermore, the detail texture reconstruction module includes residual texture learning, texture attention extraction, weak contrast detail enhancement, edge continuity constraint, detail fidelity loss construction, and detail texture reconstruction image output.

[0012] Furthermore, the color brightness adaptive correction module includes brightness equalization, color shift correction, local contrast adjustment, joint color brightness correction, and enhanced output image generation.

[0013] Furthermore, the enhancement result feedback optimization module includes noise residue assessment, detail fidelity assessment, color naturalness assessment, over-enhancement risk assessment, quality assessment result generation, feedback correction, and enhancement sample library update.

[0014] The beneficial effects achieved by adopting the above solution are as follows: (1) This invention utilizes the principle of image degradation feature analysis. Through the image degradation feature analysis module, it performs brightness distribution analysis, noise distribution analysis, blur degree analysis, contrast degradation analysis, and color shift analysis on the image to be enhanced, and extracts the image degradation feature matrix. This enables the structured representation of insufficient brightness, noise interference, blur degradation, contrast reduction, and color shift in the image to be enhanced. The technical effect is to provide a basis for degradation type and degradation area for subsequent image enhancement, and solves the problem that traditional image enhancement methods only process pixel values ​​uniformly and it is difficult to perform differentiated enhancement for different degradation areas.

[0015] (2) This invention utilizes multi-scale feature extraction of deep convolutional neural networks, constructs multi-scale enhancement branches through multi-scale depth enhancement modules, performs low-frequency brightness restoration, mid-frequency structure enhancement and high-frequency edge enhancement on the image to be enhanced, obtains low-frequency enhancement features, mid-frequency enhancement features and high-frequency enhancement features, and fuses them to generate multi-scale enhancement feature maps, realizing layered enhancement of brightness information, structural information and edge detail information. The technical effect is to improve the coordination of dark area brightness restoration, main structure enhancement and edge detail preservation, and solve the problems of insufficient dark areas, overexposure of bright areas, structural breakage and edge noise amplification that are easy to occur in traditional image enhancement algorithms.

[0016] (3) This invention utilizes residual learning, texture attention mechanism, color brightness joint correction and feedback optimization to generate a detailed texture reconstruction image through the detailed texture reconstruction module, generate an enhanced output image through the color brightness adaptive correction module, and perform feedback correction after obtaining the quality evaluation result through the enhancement result feedback optimization module. This achieves detailed reconstruction, color brightness correction and quality closed-loop optimization of the image enhancement result. The technical effect is to improve the detail fidelity, color naturalness and enhancement stability of the enhanced output image, and solve the problems of lack of result feedback, easy texture distortion, color shift and over-enhancement in traditional image enhancement processing. Attached Figure Description

[0017] Figure 1 This is a flowchart of an image enhancement processing system based on deep learning proposed in this invention. Figure 2 This is a flowchart of the image degradation feature analysis proposed in this invention. Detailed Implementation

[0018] Example 1, see Figures 1-2 The present invention provides an image enhancement processing system based on deep learning, including a degraded image acquisition module, an image degradation feature analysis module, a multi-scale depth enhancement module, a detail texture reconstruction module, a color and brightness adaptive correction module, and an enhancement result feedback optimization module; The degraded image acquisition module acquires the image to be enhanced, and collects the shooting environment data, imaging device parameters and basic image attribute data corresponding to the image to be enhanced, and sends the image to be enhanced, shooting environment data, imaging device parameters and basic image attribute data to the image degradation feature analysis module. The image degradation feature analysis module receives the image to be enhanced, shooting environment data, imaging device parameters and basic image attribute data, performs brightness distribution analysis, noise distribution analysis, blur degree analysis, contrast degradation analysis and color shift analysis on the image to be enhanced, extracts the image degradation feature matrix, and sends the image degradation feature matrix to the multi-scale depth enhancement module. The multi-scale depth enhancement module receives the image degradation feature matrix and the image to be enhanced, constructs a multi-scale enhancement branch based on a deep convolutional neural network, performs low-frequency brightness restoration, mid-frequency structure enhancement and high-frequency edge enhancement on the image to be enhanced, generates a multi-scale enhancement feature map, and sends the multi-scale enhancement feature map to the detail texture reconstruction module. The detail texture reconstruction module receives multi-scale enhanced feature maps, and based on residual learning and texture attention mechanisms, reconstructs edge contours, fine textures and weak contrast regions in the image to be enhanced, generating a detail texture reconstruction image, and sends the detail texture reconstruction image to the color brightness adaptive correction module. The color brightness adaptive correction module receives the detailed texture reconstruction image, performs brightness equalization, color shift correction and local contrast adjustment on the detailed texture reconstruction image according to the brightness distribution features and color shift features in the image degradation feature matrix, generates an enhanced output image, and sends the enhanced output image to the enhancement result feedback optimization module. The enhancement result feedback optimization module receives the enhanced output image, performs noise residue assessment, detail fidelity assessment, color naturalness assessment, and over-enhancement risk assessment on the enhanced output image to obtain a quality assessment result, and performs feedback correction on the multi-scale depth enhancement module, detail texture reconstruction module, and color brightness adaptive correction module based on the quality assessment result, ultimately realizing a closed-loop image enhancement processing that includes degraded image acquisition, degradation feature analysis, multi-scale depth enhancement, detail texture reconstruction, color brightness correction, and enhancement result feedback optimization.

[0019] Example 2: Based on all the above examples, the degraded image acquisition module acquires the image to be enhanced and collects the corresponding shooting environment data, imaging device parameters, and basic image attribute data, providing an input basis for subsequent deep learning enhancement processing; the specific operation is as follows: Image acquisition to be enhanced; the degraded image acquisition module receives at least one image to be enhanced from low-light images, foggy images, motion-blurred images, noisy images, overexposed images, and low-contrast images; for monitoring scenarios, the degraded image acquisition module acquires nighttime monitoring images, backlight monitoring images, and haze monitoring images; for industrial inspection scenarios, the degraded image acquisition module acquires surface reflection images, edge-blurred images, and weak texture images; for mobile terminal shooting scenarios, the degraded image acquisition module acquires low-light shooting images and motion-shaky shooting images. Shooting environment data acquisition; the degraded image acquisition module collects information on ambient brightness, exposure time, shooting distance, ambient humidity, ambient haze, and light source direction; if the image to be enhanced comes from a fixed camera device, the degraded image acquisition module records the camera device's installation angle and shooting area; if the image to be enhanced comes from a moving camera device, the degraded image acquisition module records the posture change information during shooting; Imaging equipment parameter acquisition; the degraded image acquisition module acquires the ISO, shutter speed, aperture parameters, focal length parameters, white balance parameters, and image compression parameters of the imaging equipment; when the ISO is high, the noise distribution in the image to be enhanced is usually enhanced; when the shutter speed is long, the probability of motion blur in the image to be enhanced increases; when the white balance parameters are abnormal, the image to be enhanced is prone to color shift; Image basic attribute data acquisition; the degraded image acquisition module acquires the image width, image height, grayscale histogram, color histogram, gradient distribution, and local contrast distribution of the image to be enhanced; if there are local dark areas in the image to be enhanced, the degraded image acquisition module marks the low-brightness areas; if there are local overexposed areas in the image to be enhanced, the degraded image acquisition module marks the high-brightness areas; Data output: The degraded image acquisition module binds the image to be enhanced, shooting environment data, imaging device parameters, and basic image attribute data according to the image number to form degraded image input data, and sends the degraded image input data to the image degradation feature analysis module. By performing the above operations, the degraded image acquisition module can provide the subsequent image degradation feature analysis module with input data that has the correlation between environment, device, and image attributes, avoiding the subsequent enhancement process relying solely on pixel values ​​and ignoring the source of image degradation.

[0020] Example 3: Based on all the above examples, the image degradation feature analysis module is used to analyze the problems of insufficient brightness, noise interference, blur degradation, decreased contrast, and color shift in the image to be enhanced, and extracts the image degradation feature matrix to provide enhancement direction constraints for the multi-scale depth enhancement module; the specific operation is as follows: Brightness distribution analysis; The image degradation feature analysis module performs brightness statistics on the image to be enhanced, calculating the global average brightness, local average brightness, and dark area ratio of the image to be enhanced; If the dark area ratio in the image to be enhanced is too high, the image degradation feature analysis module marks the image to be enhanced as a low-light degradation image; If the local average brightness is much higher than the global average brightness, the image degradation feature analysis module marks the corresponding area as an overexposed area; Noise distribution analysis; the image degradation feature analysis module calculates the noise intensity in flat areas and distinguishes between random noise, compressed noise, and high-sensitivity noise; for areas with weak texture but large pixel fluctuations, the image degradation feature analysis module increases the noise weight to prevent the multi-scale depth enhancement module from mistaking noise for real texture; Blur level analysis; The image degradation feature analysis module calculates the blur level based on the image gradient and edge diffusion width. If the gradient in the edge region decreases and the edge diffusion width increases, it is determined that the image to be enhanced has motion blur or out-of-focus blur. For images with blurred edge contours but normal color distribution, the image degradation feature analysis module increases the structure enhancement weight. Contrast degradation analysis; the image degradation feature analysis module calculates the difference between the maximum and minimum brightness values ​​within a local window to obtain local contrast features, represented as: in, Pixel position Local contrast features corresponding to local windows; It is the maximum brightness value within the local window corresponding to the pixel position (x,y); It is the minimum brightness value within the local window corresponding to the pixel position (x,y); It is a stabilizing coefficient to prevent the denominator from being zero; if the local contrast feature is lower than the preset contrast threshold, the image degradation feature analysis module will mark the corresponding local window as a weak contrast region. Color shift analysis; the image degradation feature analysis module calculates the channel mean values ​​of the image to be enhanced in the red, green, and blue channels respectively, and calculates the color shift features, expressed as: in, It is a color shift feature; It is the mean value of the red channel of the image to be enhanced; It is the mean value of the green channel of the image to be enhanced; It is the mean value of the blue channel of the image to be enhanced; It represents the difference between the mean of the red channel and the mean of the green channel; It represents the difference between the mean of the blue channel and the mean of the green channel; It represents the difference between the mean of the red channel and the mean of the blue channel; Image degradation feature matrix construction: The image degradation feature analysis module arranges the brightness distribution features, noise distribution features, blur degree features, local contrast features, and color shift features according to the image region location to construct the image degradation feature matrix, and sends the image degradation feature matrix to the multi-scale depth enhancement module.

[0021] By performing the above operations, the image degradation feature analysis module can identify the degradation type and location in the image to be enhanced, enabling the subsequent multi-scale depth enhancement module to perform differentiated enhancement according to the degradation region.

[0022] Example 4: This example is based on all the above examples. The multi-scale depth enhancement module is designed to address the problem that traditional image enhancement algorithms, which use uniform enhancement parameters, struggle to simultaneously achieve brightness restoration, structure preservation, and detail enhancement. It constructs a low-frequency brightness restoration branch, a mid-frequency structure enhancement branch, and a high-frequency edge enhancement branch based on a deep convolutional neural network to perform layered enhancement processing on the image to be enhanced. Specifically, the operation is as follows: Multi-scale input construction: The multi-scale depth enhancement module receives the image to be enhanced and the image degradation feature matrix, downsamples the image to be enhanced to different resolution image scales, and concatenates the image degradation feature matrix with the images at different scales to form multi-scale depth enhancement input; the low-resolution scale is used to extract global brightness and large-scale fog degradation information, the medium-resolution scale is used to extract body structure and region boundary information, and the high-resolution scale is used to extract edge contours and fine texture information; Low-frequency brightness recovery; the multi-scale depth enhancement module restores the brightness of dark areas, haze brightness, and underexposed areas through the low-frequency brightness recovery branch; the low-frequency brightness recovery branch uses a large receptive field convolutional layer to extract the global brightness distribution of the image and adjusts the enhancement amplitude of dark areas according to the brightness distribution characteristics; for overexposed areas, the multi-scale depth enhancement module reduces the enhancement weight of the low-frequency brightness recovery branch to prevent overexposed areas from being further brightened; Mid-frequency structure enhancement; the multi-scale depth enhancement module enhances object contours, region boundaries, and structural levels through the mid-frequency structure enhancement branch; the mid-frequency structure enhancement branch introduces local contrast features during convolution, enabling weak contrast regions to obtain higher structural enhancement intensity; when the image degradation feature matrix shows that a certain region has blurring degradation, the mid-frequency structure enhancement branch increases the structural recovery weight of the corresponding region; High-frequency edge enhancement; the multi-scale depth enhancement module enhances edge contours, fine textures, and local abrupt change regions through a high-frequency edge enhancement branch; to avoid false enhancement of high-frequency noise, the high-frequency edge enhancement branch suppresses edge enhancement intensity by combining noise distribution characteristics; when the noise distribution characteristics of a certain region are high and the gradient direction is unstable, the multi-scale depth enhancement module reduces the high-frequency edge enhancement weight of that region; Multi-scale enhancement feature fusion; the multi-scale depth enhancement module fuses the enhancement features output from the low-frequency brightness recovery branch, the mid-frequency structure enhancement branch, and the high-frequency edge enhancement branch, represented as: in, It is a multi-scale enhanced feature map It is the low-frequency enhancement feature of the low-frequency brightness recovery branch output; It is the mid-frequency enhancement characteristic of the mid-frequency structure enhancement branch output; It is the high-frequency enhancement feature of the high-frequency edge enhancement branch output; It is the weighting coefficient of the low-frequency brightness recovery branch. It is the weighting coefficient of the mid-frequency structure enhancement branch; These are the weighting coefficients for high-frequency edge enhancement branches; Degradation-guided weight calculation; the multi-scale depth enhancement module calculates the branch weights based on the brightness distribution features, blur degree features, and noise distribution features in the image degradation feature matrix, expressed as: in, , and These are the weighting coefficients for the low-frequency brightness recovery branch, the mid-frequency structure enhancement branch, and the high-frequency edge enhancement branch; It is a normalized exponential function; It is a degenerate guiding weight matrix; It is the region degradation description vector in the image degradation feature matrix; It is a degenerate guiding bias term; Multi-scale enhanced feature map output; the multi-scale depth enhancement module sends the fused multi-scale enhanced feature map to the detail texture reconstruction module, enabling the detail texture reconstruction module to perform more refined texture restoration based on the enhanced brightness, structure and edges.

[0023] Regarding parameter adjustments: Step 1: Parameter tuning of the low-frequency brightness recovery branch weights; If the enhanced output image is too dark overall, increase the weight coefficient of the low-frequency brightness recovery branch by 0.05 each time until the main structure of the dark area is visible; If overexposure diffusion occurs in the enhanced output image, the weight coefficient of the low-frequency brightness recovery branch is reduced by 0.05 each time, and the brightness suppression intensity in the color brightness adaptive correction module is increased. Step 2: Parameter tuning of branch weights for mid-frequency structure enhancement; If the object's outline is unclear, increase the weight coefficient of the mid-frequency structure enhancement branch by 0.05 each time to enhance the region boundary and main structure; If the enhanced output image shows hard edges or structural breaks, reduce the weight coefficient of the mid-frequency structure enhancement branch by 0.03 each time. Step 3: High-frequency edge enhancement branch weight parameter tuning; If detailed textures are missing, increase the weight coefficient of the high-frequency edge enhancement branch by 0.05 each time; If noise particles are amplified, the weight coefficient of the high-frequency edge enhancement branch is reduced by 0.05 each time, and the suppression strength of the noise distribution characteristics on the high-frequency edge enhancement branch is increased. Step 4: Degradation-guided weight tuning; If the enhancement effects of different degradation regions converge, then increase the learning rate of the degradation guidance weight matrix so that the branch weights can respond more significantly to the differences in regional degradation. If the enhancement intensity varies too much within the same image, the learning rate of the degradation guide weight matrix is ​​reduced, and the smoothing constraint of the branch weights in adjacent regions is increased.

[0024] By performing the above operations, this solution addresses the problems of insufficient dark areas, overexposure of bright areas, amplified edge noise, and unbalanced structural details caused by uniform enhancement in traditional image enhancement algorithms. Instead, it breaks down the image enhancement task into low-frequency brightness restoration, mid-frequency structural enhancement, and high-frequency edge enhancement through a multi-scale depth enhancement module. Furthermore, it dynamically allocates branch weights using the image degradation feature matrix, enabling different degradation regions to achieve differentiated enhancement capabilities.

[0025] Example 5: This example is based on all the above examples. The detail texture reconstruction module is designed to address the problems of missing fine textures, discontinuous edges, and unclear weak contrast details that still exist after multi-scale enhancement. It uses residual learning and texture attention mechanisms to restore textures from the multi-scale enhanced feature maps, ensuring that the enhanced image retains realistic textures while improving brightness. Specifically, the operation is as follows: Residual texture learning; the detail texture reconstruction module receives multi-scale enhanced feature maps and inputs them into the residual reconstruction network. The residual reconstruction network learns the texture residual information between the image to be enhanced and the real clear image, avoiding significant shifts in color and brightness caused by directly reconstructing the entire image; the residual texture learning process is represented as follows: in, It is a detailed texture reconstruction image; The initial enhanced image is obtained by decoding multi-scale enhanced feature maps. It is a residual reconstruction network; These are the parameters of the residual reconstruction network; It is a multi-scale enhanced feature map; It is the texture residual information learned by the residual reconstruction network from the multi-scale enhanced feature map; Texture attention extraction; the detail texture reconstruction module calculates local gradient consistency and feature response intensity to obtain a texture attention map; the texture attention map is used to distinguish between real texture regions and noise regions; for regions with consistent orientation, continuous edges and stable local contrast, the detail texture reconstruction module improves texture attention; for regions with random pixel fluctuations and unstable orientation, the detail texture reconstruction module reduces texture attention. Weak contrast detail enhancement; the detail texture reconstruction module marks areas with low local contrast features but high texture attention as weak contrast detail areas and performs residual enhancement on these areas; for light-colored text in document images, distant outlines in surveillance images, and fine scratches in industrial images, the detail texture reconstruction module improves the texture recovery intensity of weak contrast detail areas. Edge continuity constraint; the detail texture reconstruction module performs a direction consistency check on the edge contour region. If the gradient directions of adjacent edge pixels are continuous, the edge enhancement result is retained. If an isolated abrupt change occurs in the edge region, the detail texture reconstruction module reduces the residual intensity of the corresponding region to prevent isolated noise from being reconstructed as texture. Detail fidelity loss construction; the detail texture reconstruction module introduces detail fidelity loss during training to constrain the reconstructed texture to maintain consistency with the original structure, represented as: in, It's a loss of detail in preserving authenticity. It is the gradient of the image reconstructed from detailed textures; It is the gradient of a reference clear image; It is the norm difference between the gradient of the reconstructed image and the gradient of the reference sharp image; These are the texture attention loss weight coefficients; It is a texture attention map; It is element-wise multiplication; It is a detailed texture reconstruction image; It is a clear reference image; It is pixel difference guided by texture attention; The detailed texture reconstruction module outputs the image; it then sends the reconstructed image to the color and brightness adaptive correction module, which performs brightness and color correction while preserving the detailed texture.

[0026] Regarding parameter adjustments: Step 1: Adjusting parameters for residual enhancement intensity; If the detail texture reconstruction image lacks detail, increase the output strength of the residual reconstruction network by 0.05 each time; If the reconstructed image of the detailed texture appears grainy, the output intensity of the residual reconstruction network is reduced by 0.05 each time, and the suppression of random noise by the texture attention map is improved. Step 2: Texture attention threshold tuning; If the real details are not reconstructed, the texture attention threshold is reduced by 0.03 each time, so that more weak texture areas can be included in the texture reconstruction range. If a noisy region is misidentified as a textured region, the texture attention threshold is increased by 0.03 each time. Step 3: Parameter tuning to compensate for loss of detail fidelity; If the reconstructed texture is inconsistent with the original structure, the texture attention loss weight coefficient is increased by 0.05 each time. If texture restoration is too weak, reduce the texture attention loss weight coefficient and increase the residual enhancement intensity; Step 4: Parameter tuning for edge continuity constraints; If the edge breaks, increase the edge continuity constraint strength; If the edges are oversharpened, the edge continuity constraint strength is reduced, and the local smoothing strength in the color brightness adaptive correction module is increased.

[0027] By performing the above operations, this solution addresses the problems of detail distortion, texture loss, and noise enhancement that occur after brightness enhancement in traditional image enhancement algorithms. Through residual texture learning, texture attention extraction, weak contrast detail enhancement, and edge continuity constraints, the detail texture reconstruction module can restore the real texture while suppressing noise, thereby improving the detail fidelity of the enhanced image.

[0028] Example 6: This example is based on all the above examples. The color brightness adaptive correction module is used to perform brightness equalization, color shift correction, and local contrast adjustment on the detailed texture reconstructed image to avoid color distortion, local overexposure, and unnatural contrast in the image after depth enhancement. The specific operation is as follows: Brightness equalization; the color brightness adaptive correction module performs brightness equalization on the reconstructed image based on the brightness distribution characteristics in the image degradation feature matrix; for areas with insufficient enhancement in dark areas, the color brightness adaptive correction module increases local brightness; for overexposed areas, the color brightness adaptive correction module reduces brightness gain; for areas with transition between light and dark, the color brightness adaptive correction module uses smooth transition weights to avoid obvious brightness breaks. Color shift correction; the color brightness adaptive correction module performs channel gain correction on the red, green, and blue channels based on color shift characteristics, as shown below: in, This is the output image after color channel C correction; It is the channel gain coefficient corresponding to color channel c; It is the image corresponding to color channel c in the detailed texture reconstruction image; c represents the red channel, green channel, or blue channel. Local contrast adjustment; the color brightness adaptive correction module enhances weak contrast areas locally based on local contrast characteristics and suppresses enhancement in high contrast areas to prevent over-sharpening of the image; for weak contrast areas where the texture attention map shows real texture, the color brightness adaptive correction module increases local contrast; for areas with high noise distribution characteristics, the color brightness adaptive correction module reduces the intensity of local contrast enhancement. Joint color and brightness correction; the color and brightness adaptive correction module jointly constrains the brightness equalization result, color shift correction result, and local contrast adjustment result, as shown below: in, It enhances the output image; It is a pixel value truncation function; It is a detailed texture reconstruction image It is a brightness balance correction item; It is a color shift correction item. It is a local contrast adjustment option; It is the brightness balance weighting coefficient; It is the color shift correction weight coefficient; It is the local contrast adjustment weight coefficient; Enhanced output image generation; the color and brightness adaptive correction module uses the image after joint color and brightness correction as the enhanced output image, and sends the enhanced output image to the enhancement result feedback optimization module.

[0029] Regarding parameter adjustments: If the overall enhanced output image is too dark, increase the brightness equalization weight coefficient by 0.05 each time. If the output image is locally overexposed, the brightness equalization weight coefficient is reduced and the brightness suppression intensity of the overexposed area is increased. If the enhanced output image has obvious color cast, adjust the color shift correction weight coefficient and recalculate the channel gain coefficient. If the contrast of the output image is insufficient, increase the local contrast adjustment weight coefficient; If the enhanced output image is over-sharpened or noise-enhanced, the local contrast adjustment weight coefficient is reduced, and the enhancement suppression strength in noisy areas is increased.

[0030] By performing the above operations, the color brightness adaptive correction module can perform joint correction of brightness, color and contrast based on the detailed texture reconstruction image, making the enhanced output image more visually natural and reducing color distortion and local over-enhancement.

[0031] Example 7: This example is based on all the above examples. The enhancement result feedback optimization module is used to evaluate the quality of the enhanced output image and, based on the quality evaluation results, to provide feedback corrections to the multi-scale depth enhancement module, detail texture reconstruction module, and color and brightness adaptive correction module, so that the image enhancement processing system forms an adaptive closed loop; the specific operation is as follows: Noise residue assessment; the enhancement result feedback optimization module calculates the noise residue intensity in flat and dark areas; if the noise residue intensity in flat areas of the enhanced output image increases, the enhancement result feedback optimization module reduces the weight of the high-frequency edge enhancement branch and increases the texture attention threshold in the detail texture reconstruction module; Detail fidelity assessment; the enhancement result feedback optimization module calculates the detail fidelity score based on the gradient continuity and texture response consistency of the enhanced output image; if the detail fidelity score is too low, the enhancement result feedback optimization module increases the residual enhancement intensity of the detail texture reconstruction module and increases the weight of the mid-frequency structure enhancement branch; Color naturalness assessment; Enhancement result feedback optimization module calculates the channel equalization and color shift of the enhanced output image; If color naturalness is insufficient, enhancement result feedback optimization module adjusts the color shift correction weight coefficient and channel gain coefficient in the color brightness adaptive correction module; Over-enhancement risk assessment; the enhancement result feedback optimization module calculates the over-enhancement risk score based on the proportion of saturated pixels in brightness, the proportion of over-sharpened edges, and the intensity of local contrast abrupt changes, expressed as: in, It is through enhanced risk scoring; It is the brightness saturation weighting coefficient; It refers to the percentage of pixels with saturated brightness; It is the edge oversharpening weight coefficient; It is the edge oversharpening ratio; It is the weighting coefficient for local contrast abrupt changes; It is the intensity of local contrast abrupt changes; Feedback Correction: Based on the results of noise residue assessment, detail fidelity assessment, color naturalness assessment, and over-enhancement risk assessment, the enhancement result feedback optimization module corrects the parameters of the multi-scale depth enhancement module, detail texture reconstruction module, and color brightness adaptive correction module. If the over-enhancement risk score exceeds the preset risk threshold, the enhancement result feedback optimization module reduces the weight coefficients of the low-frequency brightness recovery branch, the high-frequency edge enhancement branch, and the local contrast adjustment branch. If the color naturalness score is lower than the preset naturalness threshold, the enhancement result feedback optimization module readjusts the channel gain coefficients. If the detail fidelity score is lower than the preset fidelity threshold, the enhancement result feedback optimization module increases the residual enhancement intensity of the detail texture reconstruction module. The enhancement sample library is updated; the enhancement result feedback optimization module records the image to be enhanced, the image degradation feature matrix, the multi-scale enhancement feature map, the detail texture reconstruction image, the enhancement output image, and the quality evaluation result as enhancement samples, and adds the enhancement samples to the enhancement sample library; the enhancement result feedback optimization module performs periodic training and updates on the multi-scale depth enhancement module, the detail texture reconstruction module, and the color and brightness adaptive correction module based on the enhancement sample library.

[0032] By performing the above operations, the enhancement result feedback optimization module can feed back the quality assessment results of the enhanced output image to the previous enhancement process, enabling the image enhancement processing system to continuously adjust the enhancement strategy under different degradation types, different shooting environments and different device parameters, thereby improving the stability and adaptability of the image enhancement processing system.

Claims

1. A deep learning-based image enhancement processing system, characterized in that: It includes a degraded image acquisition module, an image degradation feature analysis module, a multi-scale depth enhancement module, a detail texture reconstruction module, a color and brightness adaptive correction module, and an enhancement result feedback optimization module; The degraded image acquisition module acquires the image to be enhanced and collects the shooting environment data, imaging device parameters and basic image attribute data corresponding to the image to be enhanced, forming degraded image input data; The image degradation feature analysis module receives degraded image input data and performs brightness distribution analysis, noise distribution analysis, blur degree analysis, contrast degradation analysis, and color shift analysis on the image to be enhanced, extracting the image degradation feature matrix. The multi-scale depth enhancement module receives the image degradation feature matrix and the image to be enhanced, constructs a multi-scale enhancement branch based on a deep convolutional neural network, performs low-frequency brightness restoration, mid-frequency structure enhancement and high-frequency edge enhancement on the image to be enhanced, and obtains low-frequency enhancement features, mid-frequency enhancement features and high-frequency enhancement features, and generates a multi-scale enhancement feature map based on the low-frequency enhancement features, mid-frequency enhancement features and high-frequency enhancement features. The detail texture reconstruction module receives multi-scale enhanced feature maps and, based on residual learning and texture attention mechanisms, reconstructs edge contours, fine textures, and weak contrast regions in the image to be enhanced, generating a detail texture reconstruction image. The color brightness adaptive correction module receives the detailed texture reconstruction image and performs brightness equalization, color shift correction and local contrast adjustment on the detailed texture reconstruction image according to the brightness distribution features and color shift features in the image degradation feature matrix to generate an enhanced output image. The enhancement result feedback optimization module receives the enhanced output image, performs noise residue assessment, detail fidelity assessment, color naturalness assessment, and over-enhancement risk assessment on the enhanced output image to obtain quality assessment results, and performs feedback corrections on the multi-scale depth enhancement module, detail texture reconstruction module, and color brightness adaptive correction module based on the quality assessment results.

2. The image enhancement processing system based on deep learning according to claim 1, characterized in that: The degraded image acquisition module specifically includes: Image to be enhanced; receive at least one of the following: low-light image, foggy image, motion-blurred image, noisy image, overexposed image, and low-contrast image; Shooting environment data acquisition; collecting information on ambient brightness, exposure time, shooting distance, ambient humidity, ambient haze, and light source direction; Imaging equipment parameter acquisition; acquiring the ISO, shutter speed, aperture parameters, focal length parameters, white balance parameters, and image compression parameters of the imaging equipment; Image basic attribute data acquisition; acquisition of image width, image height, grayscale histogram, color histogram, gradient distribution, and local contrast distribution of the image to be enhanced; Data output: The image to be enhanced, shooting environment data, imaging equipment parameters, and basic image attribute data are bound according to the image number to form degraded image input data, and the degraded image input data is sent to the image degradation feature analysis module.

3. The image enhancement processing system based on deep learning according to claim 2, characterized in that: The image degradation feature analysis module specifically includes: Brightness distribution analysis; perform brightness statistics on the image to be enhanced, and calculate the global average brightness, local average brightness, and dark area ratio of the image to be enhanced; Noise distribution analysis; calculate noise intensity in flat areas and distinguish between random noise, compressed noise, and high-sensitivity noise; Blur analysis; the degree of blur is calculated based on image gradient and edge diffusion width; Contrast degradation analysis; calculate the difference between the maximum and minimum brightness values ​​within a local window to obtain local contrast characteristics; Color shift analysis: Calculate the channel mean values ​​of the image to be enhanced in the red, green, and blue channels respectively, and calculate the color shift characteristics; Image degradation feature matrix construction: The brightness distribution features, noise distribution features, blur degree features, local contrast features, and color shift features are arranged according to the image region location to construct the image degradation feature matrix, and the image degradation feature matrix is ​​sent to the multi-scale depth enhancement module.

4. The image enhancement processing system based on deep learning according to claim 3, characterized in that: The multi-scale depth enhancement module specifically includes: Multi-scale input construction: Receive the image to be enhanced and the image degradation feature matrix, downsample the image to be enhanced to different resolution image scales, and concatenate the image degradation feature matrix with the images at different scales to form a multi-scale depth enhancement input; Low-frequency brightness recovery: Based on multi-scale depth enhancement input, the brightness of dark areas, haze brightness and underexposed areas are recovered through the low-frequency brightness recovery branch to obtain low-frequency enhancement features; Mid-frequency structure enhancement: Based on multi-scale depth enhancement input, the object contour, region boundary and structural hierarchy are enhanced through the mid-frequency structure enhancement branch to obtain mid-frequency enhancement features; High-frequency edge enhancement: Based on multi-scale depth enhancement input, the edge contour, fine texture and local abrupt regions are enhanced through a high-frequency edge enhancement branch to obtain high-frequency enhancement features; Multi-scale enhancement feature fusion: Low-frequency enhancement features, mid-frequency enhancement features, and high-frequency enhancement features are fused to obtain a multi-scale enhancement feature map; Degradation-guided weight calculation: The branch weights are calculated based on the brightness distribution features, blur degree features, and noise distribution features in the image degradation feature matrix. Output multi-scale enhanced feature maps; send the multi-scale enhanced feature maps to the detail texture reconstruction module.

5. The image enhancement processing system based on deep learning according to claim 4, characterized in that: The detailed texture reconstruction module specifically includes: Residual texture learning: Receive multi-scale enhanced feature maps, input the multi-scale enhanced feature maps into the residual reconstruction network, and learn the texture residual information between the image to be enhanced and the real clear image through the residual reconstruction network; Texture attention extraction; calculate local gradient consistency and feature response intensity to obtain the texture attention map; Weak contrast detail enhancement: Based on local contrast features and texture attention maps, regions with low local contrast features and high texture attention maps are marked as weak contrast detail regions, and residual enhancement is performed on the weak contrast detail regions. Edge continuity constraints; directional consistency checks are performed on the edge contour region; Detail-preserving loss construction: Detail-preserving loss is introduced during training to constrain the reconstructed texture to maintain consistency with the original structure; Output the detailed texture reconstructed image; send the detailed texture reconstructed image to the color and brightness adaptive correction module.

6. The image enhancement processing system based on deep learning according to claim 5, characterized in that: The color brightness adaptive correction module specifically includes: Brightness equalization; based on the brightness distribution characteristics in the image degradation feature matrix, brightness equalization is performed on the reconstructed image of detailed texture; Color shift correction; channel gain correction is performed on the red, green, and blue channels based on color shift characteristics; Local contrast adjustment; based on local contrast characteristics, locally enhance weak contrast areas and suppress enhancement in high contrast areas; Joint color and brightness correction; combining the brightness equalization results, color shift correction results, and local contrast adjustment results for constraint. Enhanced output image generation: The image after joint color and brightness correction is used as the enhanced output image, and the enhanced output image is sent to the enhancement result feedback optimization module.

7. The image enhancement processing system based on deep learning according to claim 6, characterized in that: The enhanced result feedback optimization module specifically includes: Noise residual assessment; calculation of noise residual intensity in flat and dark areas; Detail fidelity assessment; Calculate detail fidelity score based on gradient continuity and texture response consistency of the enhanced output image; Color naturalness assessment; calculation of channel equalization and color shift in the enhanced output image; Over-enhancement risk assessment; the over-enhancement risk score is calculated based on the proportion of saturated pixels in brightness, the proportion of over-sharpened edges, and the intensity of local contrast abrupt changes. Quality assessment results are generated based on noise residual intensity, detail fidelity score, channel equalization, color shift, and over-enhancement risk score. Feedback and correction: Based on the quality assessment results, the parameters of the multi-scale depth enhancement module, the detail texture reconstruction module, and the color and brightness adaptive correction module are corrected. Enhancement sample library update: Record the image to be enhanced, image degradation feature matrix, multi-scale enhancement feature map, detail texture reconstruction image, enhancement output image and quality evaluation result as enhancement samples, and supplement the enhancement sample library. Based on the enhancement sample library, perform periodic training and update of multi-scale depth enhancement module, detail texture reconstruction module and color brightness adaptive correction module.

Images