An AI vision-based image enhancement method, device and medium

Abstract

The application discloses an image enhancement method and device based on AI vision and a medium, relates to the technical field of image enhancement, and comprises the following steps: generating an enhancement control strategy according to effective semantic region information and a degradation token library; controlling the first image enhancement of a to-be-enhanced image by using the enhancement control strategy; and outputting a preliminary enhanced image; performing enhancement understanding bridge analysis on the preliminary enhanced image; comparing the enhanced state of each effective semantic region with the degradation type, the degradation level and the acceptable enhancement boundary before the first image enhancement; identifying the recession state, the residual state of the original degradation and the new degradation state generated by the first image enhancement; and establishing a degradation migration relationship according to the degradation state change before and after the first image enhancement and the corresponding enhancement operation. The application reduces local over-enhancement and repeated adjustment by performing enhancement understanding bridge analysis on the preliminary enhanced image, and improves the regional consistency and controllability of the enhancement result.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to an image enhancement method, device and medium based on AI vision. Background Technology

[0002] With the development of deep learning visual perception, semantic segmentation, and adaptive image processing technologies, existing image enhancement methods have gradually evolved from traditional processing methods based on histogram equalization, filtering and sharpening, and fixed parameter mapping to intelligent enhancement methods based on convolutional neural networks, visual Transformers, and multi-scale feature fusion. These methods typically first extract visual features such as brightness, texture, edges, and local contrast, and then configure processing parameters such as exposure compensation, contrast adjustment, detail restoration, and edge enhancement according to semantic regions or degradation categories. Furthermore, they constrain the enhancement results through global quality indicators, local statistical features, or perceptual loss, thereby achieving adaptive image enhancement for different content regions.

[0003] Existing technologies typically determine whether to continue processing after an enhancement based on the overall quality score, local sharpness index, or single degradation detection result of the enhanced image. In scenarios containing multiple semantic regions with significant differences in degradation levels across different regions, the original degradation may only partially disappear, while new states such as brightness overshoot, texture oversharpening, or edge distortion may occur due to the effect of local enhancement parameters. This is because existing processing flows focus more on degradation identification and enhancement parameter generation before enhancement, and rarely establish a continuous correspondence between the degradation state before enhancement, the actual enhancement operation, and the state changes after enhancement. They also lack a mechanism for uniformly distinguishing between residual degradation and new degradation based on the acceptable enhancement boundaries of semantic regions, making it difficult to accurately determine the area of ​​action, adjustment direction, and allowable parameter range for subsequent re-enhancement. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI vision-based image enhancement method to solve the problem of difficulty in accurately identifying and directionally correcting residual degradation and new degradation after a single image enhancement.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an image enhancement method based on AI vision, comprising: acquiring an image to be enhanced, performing unified preprocessing on the image to be enhanced to generate a standardized analysis image, and simultaneously performing basic statistical analysis on the image to be enhanced to obtain basic statistical information; performing AI visual analysis on the standardized analysis image to output effective semantic region information, and determining the region degradation state in the standardized analysis image in conjunction with the basic statistical information, determining the corresponding acceptable enhancement boundary based on the region degradation state, and obtaining a degradation representation token library; generating an enhancement control strategy based on the effective semantic region information and the degradation representation token library, using the enhancement control strategy to control the first image enhancement of the image to be enhanced, and outputting a preliminary enhanced image; and performing AI visual analysis on the preliminary enhanced image to obtain basic statistical information. The enhanced image undergoes bridging analysis to compare the enhanced state of each effective semantic region with the degradation type, degradation level, and acceptable enhancement boundary before the first image enhancement. This identifies the original degradation fading state, residual state, and newly generated degradation state produced by the first image enhancement. Degradation migration relationships are established based on the changes in degradation state before and after the first image enhancement and the corresponding enhancement operations. Targeted re-enhancement is then performed on the initially enhanced image. Based on the degradation migration relationship, the regions to be re-enhanced and the enhancement direction of the initially enhanced image are determined. The enhancement influence relationship between the regions to be re-enhanced and adjacent effective semantic regions is determined by combining spatial correlation. The allowable adjustment range of local enhancement parameters is determined based on the acceptable enhancement boundary to obtain the final enhanced image.

[0007] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the steps of acquiring the image to be enhanced, performing unified preprocessing on the image to be enhanced to generate a standardized analysis image, and performing basic statistical analysis on the image to be enhanced to obtain basic statistical information are as follows: The image to be enhanced is obtained, and its size is standardized and its color space is unified to generate a unified image to be enhanced. The unified image to be enhanced is then subjected to brightness extraction and pixel normalization, and preliminary suppression is performed to generate a standardized analysis image. Basic statistical features are extracted from the image to be enhanced through basic statistical analysis, and then normalized and sequentially stitched to generate basic statistical information.

[0008] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the specific steps of performing AI visual analysis on the standardized analysis image and outputting effective semantic region information are as follows: The standardized analysis image is input into the AI ​​visual analysis model, which performs multi-layer visual feature extraction, multi-scale feature mapping and cross-layer fusion processing on the standardized analysis image to generate a semantic analysis feature map. Candidate regions are extracted based on semantic parsing feature maps to generate candidate semantic regions. Then, semantic category discrimination, region boundary localization, and recognition confidence calculation are performed on the candidate semantic regions to select effective semantic regions. Based on the semantic category labels, regional location parameters, and recognition confidence of the effective semantic region, spatial associations and semantic priority relationships are established and encapsulated to form effective semantic region information.

[0009] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the steps of determining the degradation state of regions in the standardized analysis image by combining basic statistical information, determining the corresponding acceptable enhancement boundaries based on the degradation state, and obtaining a degradation representation token library are as follows: Based on the effective semantic region information, each effective semantic region is located, and the brightness distribution, texture clarity, edge continuity and local contrast changes are extracted. These are then compared with basic statistical information to determine the degradation type and degradation level of each effective semantic region. Based on the semantic category label, degradation type, and degradation level of the effective semantic region, the allowable range of variation corresponding to brightness, contrast, texture, and edge state is determined to obtain the acceptable enhancement boundary; The degradation type, degradation level, and acceptable enhancement boundary of the effective semantic region are associated with the corresponding region location and encoded to generate regional degradation representation tokens. The regional degradation representation tokens are then organized according to the correspondence between effective semantic regions to form a degradation representation token library.

[0010] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the specific steps for generating an enhancement control strategy based on effective semantic region information and a degradation representation token library are as follows: The effective semantic region information is associated with the degenerate representation token library to output the target enhancement mode and local enhancement parameters; Based on the semantic category labels, degradation types, degradation levels, and acceptable enhancement boundaries of the effective semantic region, the target enhancement mode and local enhancement parameters are configured to generate candidate enhancement configurations. Then, based on spatial correlation and semantic priority, the candidate enhancement configurations are reconciled and ordered to form an enhancement control strategy.

[0011] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the step of using an enhancement control strategy to control the first image enhancement of the image to be enhanced and outputting a preliminary enhanced image includes the following specific steps: The image to be enhanced is divided into multiple enhancement regions. According to the target enhancement mode and local enhancement parameters corresponding to each enhancement region, the image to be enhanced is subjected to region adaptive enhancement processing based on the enhancement control strategy to obtain the region enhancement result. The enhancement transition between adjacent regions is coordinated to form transition coordination information, which is then fused to output a preliminary enhanced image.

[0012] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the steps of identifying the original degradation state, the residual state, and the newly generated degradation state produced by the first image enhancement, and establishing a degradation migration relationship based on the changes in degradation state before and after the first image enhancement and the corresponding enhancement operation, are as follows: The brightness, texture, edge and local contrast status of each effective semantic region in the preliminary enhanced image are extracted and compared with the corresponding degradation type and degradation level before the first image enhancement. The original degradation fading status and residual status are determined based on the changes in degradation characterization. The enhanced state is compared with the corresponding acceptable enhancement boundary. When the state that did not meet the degradation judgment condition before the first image enhancement exceeds the corresponding acceptable enhancement boundary after enhancement, it is determined to be a newly degraded state, and the corresponding newly degraded type and state change degree are recorded. Based on the changes in degradation state before and after the first image enhancement of each effective semantic region, the corresponding target enhancement mode and local enhancement parameters are associated to determine the correspondence between the original degradation state, residual state and newly generated degradation state after the corresponding enhancement operation, thus forming a degradation migration relationship.

[0013] As a preferred embodiment of the AI ​​vision-based image enhancement method of the present invention, the method involves performing targeted re-enhancement on the initially enhanced image, determining the region to be re-enhanced and the enhancement direction of the initially enhanced image based on the degradation migration relationship, determining the enhancement influence relationship between the region to be re-enhanced and adjacent effective semantic regions based on spatial correlation, and determining the allowable adjustment range of local enhancement parameters based on acceptable enhancement boundaries to obtain the final enhanced image. The specific steps are as follows: The regions corresponding to the residual state and the newly degraded state are identified as regions to be re-enhanced. The original enhancement direction corresponding to the residual state is determined according to the degradation migration relationship, and the reverse suppression direction is determined according to the enhancement operation corresponding to the newly degraded state, thus obtaining the enhancement direction of each region to be re-enhanced. Obtain the remaining change between the current enhancement state of the region to be enhanced and the corresponding acceptable enhancement boundary, and determine the effective semantic region adjacent to the region to be enhanced based on the spatial correlation. Obtain the remaining change, determine the allowable adjustment range of the local enhancement parameters based on the remaining change, and correct the local enhancement parameters that exceed the allowable adjustment range. Directional re-enhancement is performed on each region to be re-enhanced according to the enhancement direction and the corrected local enhancement parameters, and the regions that have been re-enhanced are then fused to obtain the final enhanced image.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, the memory storing a computer program, wherein the computer program, when executed by the processor, implements any step of the AI ​​vision-based image enhancement method as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the AI ​​vision-based image enhancement method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by performing enhancement understanding bridging analysis on the preliminary enhanced image, the enhanced state of each effective semantic region is associated with the degradation type, degradation level, acceptable enhancement boundary and corresponding enhancement operation before enhancement, forming a degradation migration relationship. This enables unified identification of the original degradation state, residual state and newly generated degradation state, and determines the region to be re-enhanced, enhancement direction and allowable adjustment range of local parameters. This allows subsequent image enhancement to implement targeted correction for state changes, reducing local over-enhancing and repeated adjustment, and improving the regional consistency and controllability of the enhancement results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI vision-based image enhancement method.

[0019] Figure 2 This is a flowchart of semantic parsing and degenerate representation.

[0020] Figure 3 A flowchart for enhanced control and the first image enhancement.

[0021] Figure 4 A flowchart to enhance understanding of bridging analysis and directed re-enhancement. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an image enhancement method based on AI vision, including the following steps: S1. Obtain the image to be enhanced, perform unified preprocessing on the image to be enhanced to generate a standardized analysis image, and perform basic statistical analysis on the image to be enhanced to obtain basic statistical information.

[0026] S1.1. Obtain the image to be enhanced, perform size standardization and color space unification on the image to be enhanced to generate a unified image to be enhanced, perform brightness extraction and pixel normalization on the unified image to be enhanced, and perform preliminary suppression to generate a standardized analysis image.

[0027] Specifically, the process involves acquiring the image to be enhanced, standardizing its size to a uniform resolution, and converting it to a uniform color representation to generate a unified image to be enhanced. Brightness information representing the distribution of light and dark areas is then extracted from this unified image, and pixel values ​​are mapped to a uniform range, completing brightness extraction and pixel normalization. Preliminary suppression processing is then applied to the unified image, including median filtering or bilateral filtering for high-frequency noise, bright spot threshold detection combined with neighboring pixel replacement for isolated anomalous bright spots, and amplitude limiting suppression for local abrupt changes to constrain the amplitude of abrupt changes between adjacent pixels, resulting in a standardized analysis image.

[0028] S1.2. Extract basic statistical features from the image to be enhanced through basic statistical analysis, and perform normalization and sequential stitching to generate basic statistical information.

[0029] Specifically, based on the image to be enhanced and the standardized analysis image, basic statistical features of the image to be enhanced are extracted through basic statistical analysis. These basic statistical features include brightness statistics reflecting the overall brightness distribution, grayscale fluctuation statistics reflecting the degree of pixel dispersion, color statistics reflecting the color distribution state, and texture statistics reflecting the degree of texture undulation, including local entropy, gradient mean, gradient energy, and local contrast. Then, each basic statistical feature is normalized to eliminate the dimensional differences between different statistics. Finally, the normalized basic statistical features are sequentially concatenated to generate basic statistical information.

[0030] S2. Perform AI visual analysis on the standardized analysis image, output effective semantic region information, and combine basic statistical information to determine the regional degradation state in the standardized analysis image. Based on the regional degradation state, determine the corresponding acceptable enhancement boundary and obtain the degradation representation token library.

[0031] S2.1. Input the standardized analysis image into the AI ​​visual analysis model, perform multi-layer visual feature extraction, multi-scale feature mapping and cross-layer fusion processing on the standardized analysis image to generate a semantic analysis feature map.

[0032] Specifically, the standardized analysis image is input into the AI ​​visual analysis model. The AI ​​visual analysis model performs multi-layer visual feature extraction on the standardized analysis image. Edge texture information is extracted through layer-by-layer convolution, activation, and downsampling operations. Edge texture information is used to characterize the edge contours, texture details, fine-grained surface changes, and local structural information formed by grayscale changes, color changes, or local high-frequency fluctuations in the image. Local structural information is used to characterize the spatial combination relationship, geometric distribution relationship, regional organization morphology, and high-level semantic information between pixels, edges, or texture responses in local areas of the image. High-level semantic information is used to characterize the content category, semantic attribution, target attributes, and functional role of the image region in the overall visual content. Multi-scale feature mapping is performed on visual features at different levels to convert the regional responses corresponding to different receptive fields into a unified expression form. Then, the multi-scale feature mapping results are fused across layers. Through upsampling, stitching, and weighted integration, shallow detail information and deep semantic information are combined to generate a semantic analysis feature map.

[0033] It should be noted that the AI ​​visual analysis model is a pre-trained semantic segmentation model. It collects training images with labeled semantic regions, performs size and color space standardization on these images to ensure consistency with the input format of the standardized analysis images. The AI ​​visual analysis model then inputs these labeled training images into the model, outputting corresponding semantic analysis results. These results are then compared pixel-by-pixel with the corresponding semantic region labeling results in the training images to obtain semantic category bias, region boundary bias, and recognition confidence bias. A loss value is calculated based on these biases, and the network parameters in the AI ​​visual analysis model are updated accordingly. This process of inputting training images, outputting semantic analysis results, calculating loss values, and updating network parameters is repeated in multiple rounds until the semantic analysis results on the verification images stabilize, resulting in an AI visual analysis model capable of outputting semantic analysis feature maps.

[0034] S2.2. Based on the semantic parsing feature map, candidate regions are extracted to generate candidate semantic regions. Then, the candidate semantic regions are subjected to semantic category discrimination, region boundary localization and recognition confidence calculation to select effective semantic regions.

[0035] Specifically, the pixel response intensity, semantic aggregation distribution, and regional continuity in the semantic parsing feature map are detected, and regions with concentrated and spatially continuous responses are classified as candidate semantic regions. Semantic category discrimination is performed on the candidate semantic regions by comparing their category response results in the semantic parsing feature map to determine the corresponding semantic category label. Then, the region boundaries of the candidate semantic regions are located by extracting the edge change positions and region closure ranges in the semantic parsing feature map to determine the region boundaries. Simultaneously, recognition confidence is calculated for the candidate semantic regions by statistically analyzing the consistency of semantic category responses within the candidate semantic regions and the concentration of boundary localization. Finally, the semantic category discrimination results, region boundary localization results, and recognition confidence calculation results are combined to select effective semantic regions.

[0036] It should be noted that, assuming the first The candidate semantic regions are Pixel positions in semantic parsing feature maps semantic categories The response value ;No. The semantic category labels corresponding to the candidate semantic regions are: ; The consistency of semantic category responses within a candidate semantic region can then be expressed as: ; in, Indicates the first The degree of consistency of semantic category responses within each candidate semantic region; The larger the value, the more consistent the semantic category responses within the candidate semantic region; Indicates the number of the candidate semantic region; Indicates the first One candidate semantic region; Indicates the first Pixel position coordinates in candidate semantic regions; Indicates pixel position For the Semantic category labels corresponding to each candidate semantic region The response value; Indicates the first Semantic category labels corresponding to each candidate semantic region; The category number representing the semantic category; Indicates the total number of semantic categories; Indicates the first The number of pixels within each candidate semantic region; Indicates pixel position The sum of response values ​​for all semantic categories; Category identifier representing semantic category.

[0037] The degree of concentration of location at the regional boundary can be expressed as: ; in, Indicates the first The degree of concentration of the regional boundary localization of each candidate semantic region; Indicates the first The set of boundary pixels of candidate semantic regions; Indicates pixel position The boundary response value at that point.

[0038] S2.3. Based on the semantic category labels, regional location parameters, and recognition confidence of the effective semantic region, establish spatial association and semantic priority relationship, and encapsulate them to form effective semantic region information.

[0039] Specifically, spatial relationships are established by calculating the relative positions, adjacency states, overlap states, and distance distributions between effective semantic regions based on regional location parameters. Relative positions include top-bottom, left-right, and containment relationships; adjacency states characterize whether region boundaries touch; overlap states characterize whether region ranges overlap; and distance distribution characterizes the distance between region centers or the shortest boundary. When establishing semantic priority relationships, the content attributes corresponding to semantic category labels are jointly analyzed with recognition confidence. Content attributes include at least text regions, face regions, target regions, and background regions. Text regions and face regions have higher priority than target regions, and target regions have higher priority than background regions. Under the same content attributes, effective semantic regions with higher recognition confidence have higher priority than those with lower recognition confidence. When content attributes and recognition confidence are similar, the region area ratio or the degree of visual centering of the region is used as a further ranking criterion, thus forming semantic priority relationships. The semantic category labels, regional location parameters, recognition confidence, spatial relationships, and semantic priority relationships of effective semantic regions are encapsulated according to a unified field order to form effective semantic region information.

[0040] S2.4. Locate each effective semantic region based on the effective semantic region information, extract brightness distribution, texture clarity, edge continuity and local contrast changes, and compare them with basic statistical information to determine the degradation type and degradation level of each effective semantic region.

[0041] Specifically, based on the regional location parameters in the effective semantic region information, each effective semantic region is extracted from the standardized analysis image. The pixel brightness distribution and brightness variation range within the region are statistically analyzed. Local texture response is extracted and texture clarity is determined. The edge continuity is determined based on the connection state between pixels at the region edge, and local contrast changes are determined based on the brightness difference within the region. The brightness distribution, texture clarity, edge continuity, and local contrast variation of each effective semantic region are compared with the corresponding baseline state in the basic statistical information. The degradation type of the region is determined according to the feature type that deviates from the baseline state, and the degradation level is determined according to the degree of deviation of the corresponding feature from the baseline state and the distribution range of the deviated region in the effective semantic region, so as to obtain the region degradation state corresponding to each effective semantic region.

[0042] S2.5. Based on the semantic category label, degradation type, and degradation level of the effective semantic region, determine the allowable range of variation for brightness, contrast, texture, and edge state to obtain the acceptable enhancement boundary.

[0043] Specifically, the brightness, contrast, texture and edge state benchmarks of the corresponding semantic content under normal identifiable state are determined based on the semantic category labels of each effective semantic region, and the initial allowable range of variation is determined with the corresponding state benchmark as the center. The state dimensions that need to be adjusted are determined based on the degradation type, where brightness degradation corresponds to brightness state, contrast degradation corresponds to contrast state, texture degradation corresponds to texture state, and edge degradation corresponds to edge state. Then, the initial allowable variation range of the corresponding state dimensions is modified according to the degradation level. The higher the degradation level, the greater the allowable variation in the direction of degradation improvement, while limiting the variation in the direction of degradation aggravation. The upper and lower limits of allowable variation for brightness, contrast, texture, and edge states are obtained respectively. The upper and lower limits of allowable variation for each state are combined to form the acceptable enhancement boundary of each effective semantic region.

[0044] S2.6. Associate and encode the degradation type, degradation level, and acceptable enhancement boundary of the effective semantic region with the corresponding region location to generate a region degradation representation token. Organize the region degradation representation tokens according to the correspondence between effective semantic regions to form a degradation representation token library.

[0045] Specifically, region identifiers are generated based on the regional location parameters of each effective semantic region. The corresponding degradation type, degradation level, and the upper and lower limits of allowable changes in brightness, contrast, texture, and edge state are encoded separately. These are then associated and combined in the order of region identifier, degradation type, degradation level, and acceptable enhancement boundary to generate a region degradation representation token corresponding to a single effective semantic region. Using the region identifier as an index, the region degradation representation tokens are mapped to the spatial association and semantic priority relationships in the effective semantic region information. This ensures that each region degradation representation token maintains its association with its corresponding effective semantic region and adjacent effective semantic regions. The tokens are then arranged according to semantic priority to form a degradation representation token library.

[0046] S3. Based on the effective semantic region information and the degradation representation token library, generate an enhancement control strategy, use the enhancement control strategy to control the first image enhancement of the image to be enhanced, and output the preliminary enhanced image.

[0047] S3.1. Associate the effective semantic region information with the degraded representation token library, and output the target enhancement mode and local enhancement parameters.

[0048] Specifically, based on the semantic category labels, region location parameters, and recognition confidence in the effective semantic region information, the degradation representation tokens matching each effective semantic region are retrieved from the degradation representation token library. The effective semantic regions are then associated by the corresponding degradation type, degradation level, acceptable enhancement boundary, and secondary enhancement priority based on the semantic category labels. Based on the association results, the applicable enhancement direction for each effective semantic region is determined, and processing methods biased towards brightness compensation, contrast enhancement, texture enhancement, edge repair, or noise suppression are identified as target enhancement modes. Then, combined with the acceptable enhancement boundary and degradation level, the brightness adjustment amplitude, contrast adjustment amplitude, texture enhancement intensity, edge enhancement intensity, and suppression intensity under the target enhancement mode are quantified and configured, and local enhancement parameters are output.

[0049] S3.2. Based on the semantic category labels, degradation types, degradation levels and acceptable enhancement boundaries of the effective semantic region, configure the target enhancement mode and local enhancement parameters to generate candidate enhancement configurations. Based on spatial association and semantic priority, coordinate the conflicts and arrange the order of the candidate enhancement configurations to form an enhancement control strategy.

[0050] Specifically, the target enhancement mode is configured by determining the processing focus based on the content attributes and degradation state corresponding to the effective semantic region, and limiting the local enhancement parameters to the acceptable range of variation corresponding to the enhancement boundary. Then, the local enhancement parameters are refined to generate candidate enhancement configurations, determining the brightness adjustment amplitude, contrast adjustment amplitude, texture enhancement intensity, edge enhancement intensity, and noise suppression intensity based on the degradation level. Next, conflict resolution is performed based on spatial relationships, comparing local enhancement parameters that differ between adjacent and overlapping effective semantic regions, and addressing boundary adjacency positions. The local enhancement parameters are smoothly adjusted; the regions are arranged sequentially according to semantic priority, prioritizing effective semantic regions with higher semantic importance and stronger degradation impact. For example, in the same image to be enhanced, when both text and background regions are degraded, enhancement is prioritized for the text region; when both face and ordinary texture regions are degraded, enhancement is prioritized for the face region; when the semantic importance of two effective semantic regions is the same or close, enhancement is prioritized for the effective semantic region with higher degradation level and more obvious impact on content recognizability, thus forming an enhancement control strategy.

[0051] S3.3. According to the enhancement control strategy, the image to be enhanced is divided into multiple enhancement regions. Based on the target enhancement mode and local enhancement parameters corresponding to each enhancement region, the image to be enhanced is subjected to region adaptive enhancement processing according to the enhancement control strategy to obtain the region enhancement result.

[0052] Specifically, based on the region location parameters, target enhancement mode, and local enhancement parameters corresponding to each effective semantic region, the image to be enhanced is divided into multiple enhancement regions, so that each enhancement region corresponds to different semantic content and different degradation states. According to the target enhancement mode and local enhancement parameters corresponding to each of the multiple enhancement regions, adaptive region enhancement processing is performed on the image to be enhanced, and brightness adjustment, contrast adjustment, texture enhancement, edge enhancement, or noise suppression are applied to each of the multiple enhancement regions, so that each enhancement region completes targeted enhancement in its corresponding enhancement direction. The region enhancement result after adaptive region enhancement processing is output.

[0053] S3.4. Perform coordination processing on the enhancement transition between adjacent regions to form transition coordination information and fuse them to output a preliminary enhanced image.

[0054] Specifically, based on the brightness adjustment, contrast adjustment, texture enhancement, edge enhancement, and noise suppression results at the boundary junctions of adjacent regions, enhancement differences on both sides of the boundary are identified. Then, pixel changes within the boundary transition range are smoothed, allowing adjacent enhancement areas to gradually transition from one side to the other in terms of brightness, contrast, and texture changes, reducing abrupt boundary changes, brightness-dark discontinuities, and texture fragmentation. The enhancement areas after the enhancement transition coordination process are then fused, and each enhancement area is backfilled into its corresponding region in the image to be enhanced according to its original spatial location. For pixels within the boundary transition range, a weighted fusion method based on boundary distance is used for smooth stitching. For overlapping regions, fusion weights are assigned according to the semantic priority relationship of the corresponding enhancement areas, and then weighted fusion is performed to output a preliminary enhanced image.

[0055] S4. Perform enhancement understanding bridging analysis on the preliminary enhanced image, compare the enhanced state of each effective semantic region with the degradation type, degradation level and acceptable enhancement boundary before the first image enhancement, identify the original degradation fading state, residual state and newly generated degradation state generated by the first image enhancement, and establish degradation transfer relationship based on the degradation state changes before and after the first image enhancement and the corresponding enhancement operation.

[0056] S4.1. Extract the brightness, texture, edge and local contrast status of each effective semantic region in the preliminary enhanced image, compare it with the corresponding degradation type and degradation level before the first image enhancement, and determine the original degradation fading state and residual state based on the changes in degradation characterization.

[0057] Specifically, based on the regional location parameters in the effective semantic region information, each effective semantic region is located in the preliminary enhanced image. The brightness distribution, texture clarity, edge continuity, and local contrast changes of each effective semantic region are obtained to form the enhanced state of each effective semantic region. The degradation type and degradation level of each effective semantic region before the first image enhancement are read from the degradation characterization token library. Based on the degradation type, the enhanced state to be compared is determined, and the enhanced state is compared with the corresponding baseline state in the basic statistical information. When the enhanced state recovers to the corresponding reference state range, the original degradation is determined as the fading state; when there is still a deviation in the same direction as before the first image enhancement between the enhanced state and the corresponding reference state, the original degradation is determined as the residual state. The case where the degree of deviation is reduced but still has not recovered to the corresponding reference state range is still determined as the residual state. Thus, the fading state and residual state of the original degradation are determined according to the changes in degradation characterization.

[0058] S4.2. Compare the enhanced state with the corresponding acceptable enhancement boundary. When the state that did not meet the degradation judgment condition before the first image enhancement exceeds the corresponding acceptable enhancement boundary after enhancement, it is determined to be a newly degraded state, and the corresponding newly degraded type and state change degree are recorded.

[0059] Specifically, based on the regional location parameters in the effective semantic region information, each effective semantic region is located in the preliminary enhanced image, and the brightness distribution, texture clarity, edge continuity and local contrast changes of each effective semantic region before the first image enhancement are read, and the enhanced state of the corresponding effective semantic region in the preliminary enhanced image is obtained. For images that are within the baseline state range corresponding to the basic statistical information and have not formed a corresponding degradation type before the first image enhancement, the enhanced brightness state, texture state, edge state, and local contrast state are compared with the acceptable enhancement boundaries of the corresponding effective semantic regions in the degradation representation token library. When the enhanced state is still within the corresponding acceptable enhancement boundary, the corresponding state remains unchanged. When the enhanced state exceeds the upper or lower limit of the allowable change of the corresponding acceptable enhancement boundary, the new degradation type is determined according to the state type and change direction of the boundary crossing. Among them, brightness state boundary crossing corresponds to brightness-type new degradation, texture state boundary crossing corresponds to texture-type new degradation, edge state boundary crossing corresponds to edge-type new degradation, and local contrast state boundary crossing corresponds to contrast-type new degradation. Then, the degree of state change is determined according to the degree to which the enhanced state deviates from the corresponding acceptable enhancement boundary, thus obtaining the new degradation state, and recording the corresponding new degradation type and state change degree.

[0060] S4.3. Based on the changes in degradation state before and after the first image enhancement of each effective semantic region, associate the corresponding target enhancement mode and local enhancement parameters, determine the correspondence between the original degradation state, residual state and newly generated degradation state after the corresponding enhancement operation, and form a degradation migration relationship.

[0061] Specifically, based on the regional location parameters of the effective semantic region, the degradation type and degradation level before the first image enhancement, the target enhancement mode and local enhancement parameters used in the first image enhancement process, and the fading state, residual state and newly generated degradation state after the first image enhancement are matched according to the same effective semantic region. For fading states, the original degradation type and degradation level are associated with the target enhancement mode and local enhancement parameters that restore the corresponding degradation representation to the baseline state range corresponding to the basic statistical information. For residual states, the original degradation type and degradation level are associated with the degradation representation that still maintains the same deviation direction after enhancement, the corresponding target enhancement mode and local enhancement parameters, and the residual changes of the original degradation after the corresponding enhancement operation are determined based on the changes in degradation level before and after the first image enhancement. For newly nascent degradation states, the target enhancement mode and local enhancement parameters that affect the same state are found based on the brightness, texture, edge or local contrast state corresponding to the newly nascent degradation type, and the association between the corresponding enhancement operation and the newly nascent degradation state is determined based on the direction of action of the local enhancement parameters and the direction of change of the newly nascent degradation state. The degeneration migration relationship is formed by organizing the original degeneration type, degeneration level, corresponding enhancement operation, and the correspondence between the fading state, residual state and newly degeneration state in each effective semantic region.

[0062] S5. Perform targeted re-enhancement on the preliminary enhanced image. Determine the region to be re-enhanced and the enhancement direction of the preliminary enhanced image based on the degradation migration relationship. Combine the spatial correlation relationship to determine the enhancement influence relationship between the region to be re-enhanced and the adjacent effective semantic region. Determine the allowable adjustment range of the local enhancement parameters based on the acceptable enhancement boundary to obtain the final enhanced image.

[0063] S5.1. The regions corresponding to the residual state and the regions corresponding to the newly degenerated state are identified as regions to be re-enhanced. The original enhancement direction corresponding to the residual state is determined according to the degradation migration relationship, and the reverse suppression direction is determined according to the enhancement operation corresponding to the newly degenerated state, thus obtaining the enhancement direction of each region to be re-enhanced.

[0064] Specifically, based on the degradation migration relationship, the fading state, residual state and newly degraded state of each effective semantic region after the first image enhancement are read, and the effective semantic regions with residual state and newly degraded state are identified as regions to be re-enhanced; For the regions to be enhanced corresponding to the residual state, the original degradation type, the target enhancement mode used in the first image enhancement, and the local enhancement parameters are read from the degradation migration relationship. Based on the change in degradation level before and after the first image enhancement, it is determined whether the corresponding enhancement operation reduces the original degradation. When the original degradation is reduced, the direction that promotes the original degradation to continue to recover to the baseline state corresponding to the basic statistical information is determined as the original enhancement direction. For the regions to be enhanced corresponding to the newly degraded state, the newly degraded type, the degree of state change, and the target enhancement mode and local enhancement parameters associated with the newly degraded state are read from the degradation migration relationship. Based on the out-of-bounds direction of the brightness, texture, edge, or local contrast state corresponding to the newly degraded state, the reverse suppression direction opposite to the direction of the local enhancement parameters that caused the out-of-bounds is determined, thereby obtaining the enhancement direction of each region to be enhanced.

[0065] S5.2. Obtain the remaining change between the current enhancement state of the region to be enhanced and the corresponding acceptable enhancement boundary, and determine the effective semantic region adjacent to the region to be enhanced based on the spatial correlation. Obtain the remaining change, determine the allowable adjustment range of the local enhancement parameters based on the remaining change, and correct the local enhancement parameters that exceed the allowable adjustment range.

[0066] Specifically, based on the enhancement direction of each region to be re-enhanced, the current enhancement state of each region to be re-enhanced and the corresponding acceptable enhancement boundary in the degradation characterization token library are read; for the region to be re-enhanced corresponding to the residual state, the upper limit or lower limit of the allowable change corresponding to the current enhancement state is determined along the original enhancement direction, and the state changes that can still occur between the current enhancement state and the corresponding upper limit or lower limit of the allowable change are determined as the remaining amount of change of the region to be re-enhanced. For the region to be re-enhanced corresponding to the nascent degenerate state, determine the upper or lower limit of allowable change that the current enhancement state has already exceeded along the reverse inhibition direction, and determine the state change required to restore the current enhancement state to the corresponding upper or lower limit of allowable change as the remaining change amount of the region to be re-enhanced. Based on the spatial correlation in the effective semantic region information, adjacent effective semantic regions that border the boundary of the region to be re-enhanced and overlap with the enhancement transition range are identified. The current enhancement state and corresponding acceptable enhancement boundary of the adjacent effective semantic regions are read. According to the brightness, contrast, texture and edge state of the local enhancement parameters of the region to be re-enhanced, the states that may be affected by the corresponding local enhancement parameters in the adjacent effective semantic regions are determined. The state changes that can still be tolerated between the current enhancement state of the adjacent effective semantic regions and the corresponding upper or lower limit of allowable changes are determined as the remaining change amount of the adjacent effective semantic regions. Combining the remaining change amount of the region to be re-enhanced and the remaining change amount of the adjacent effective semantic regions, the remaining change amount with stricter constraints on the current enhancement direction is used as the adjustment constraint of the local enhancement parameters. Then, the allowable adjustment range of the local enhancement parameters is calculated according to the response relationship between the local enhancement parameters used in the first image enhancement in the degradation migration relationship and the corresponding state changes. The proposed adjustment range of the local enhancement parameters is compared with the allowable adjustment range. If the adjustment range of the local enhancement parameters does not exceed the allowable adjustment range, the adjustment range of the local enhancement parameters is retained. If the adjustment range of the local enhancement parameters exceeds the allowable adjustment range, the adjustment range of the local enhancement parameters is corrected to the allowable adjustment range.

[0067] S5.3. Perform directional re-enhancement on each region to be re-enhanced according to the enhancement direction and the corrected local enhancement parameters, and fuse the regions that have been re-enhanced to obtain the final enhanced image.

[0068] Specifically, based on the regional location parameters of each region to be enhanced, the corresponding image regions are extracted from the preliminary enhanced image, and the enhancement direction and corrected local enhancement parameters of each region to be enhanced are read. For the regions to be enhanced corresponding to the residual state, brightness enhancement, contrast enhancement, texture enhancement, or edge enhancement corresponding to the original degradation type are performed according to the original enhancement direction, and the degree of enhancement is controlled according to the corrected local enhancement parameters, so that the enhanced state corresponding to the residual degradation is restored to the baseline state corresponding to the basic statistical information. For the newly degraded regions corresponding to the areas to be enhanced, reverse enhancement processing is performed on the brightness, texture, edge, or local contrast states that exceed the limits according to the reverse suppression direction. The degree of reverse enhancement is controlled according to the corrected local enhancement parameters to restore the newly degraded states to the corresponding acceptable enhancement boundaries. After completing the directional enhancement of each region to be enhanced, the image content in the initial enhanced image that does not belong to the region to be enhanced remains unchanged. The connection boundary between the region to be enhanced and the surrounding image region is determined according to the regional position parameters of the region to be enhanced. A continuous transition range is set on both sides of the connection boundary. The degree of fusion between the content of the enhanced image and the content of the initial enhanced image is gradually changed according to the distance from the pixel position to the connection boundary, so that the brightness, texture, edge, and local contrast states on both sides of the connection boundary transition continuously, completing the fusion of the enhanced region and the initial enhanced image, and obtaining the final enhanced image.

[0069] This embodiment also provides a computer device applicable to the AI ​​vision-based image enhancement method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI ​​vision-based image enhancement method proposed in the above embodiment.

[0070] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0071] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the AI ​​vision-based image enhancement method as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0072] In summary, this invention achieves unified identification of the original degradation state, residual state, and newly generated degradation state by performing enhancement understanding bridging analysis on the preliminary enhanced image, associating the post-enhancement state of each effective semantic region with the pre-enhancement degradation type, degradation level, acceptable enhancement boundary, and corresponding enhancement operation, thus forming a degradation migration relationship. Based on this, it determines the region to be re-enhanced, the enhancement direction, and the allowable adjustment range of local parameters, enabling subsequent image enhancement to implement targeted corrections for state changes, reducing local over-enhancement and redundant adjustments, and improving the regional consistency and controllability of the enhancement results.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An image enhancement method based on AI vision, characterized in that, include: The image to be enhanced is acquired and uniformly preprocessed to generate a standardized analysis image. At the same time, basic statistical analysis is performed on the image to be enhanced to obtain basic statistical information. AI visual analysis is performed on the standardized analysis image to output effective semantic region information. Combined with basic statistical information, the degradation state of the region in the standardized analysis image is determined. Based on the degradation state of the region, the corresponding acceptable enhancement boundary is determined, and a degradation representation token library is obtained. Based on the effective semantic region information and the degradation representation token library, an enhancement control strategy is generated. The enhancement control strategy is used to control the first image enhancement of the image to be enhanced, and the preliminary enhanced image is output. Enhancement understanding bridging analysis is performed on the initially enhanced image. The enhanced state of each effective semantic region is compared with the degradation type, degradation level and acceptable enhancement boundary before the first image enhancement. The original degradation fading state, residual state and newly generated degradation state generated by the first image enhancement are identified. Degradation transfer relationship is established based on the degradation state changes before and after the first image enhancement and the corresponding enhancement operation. The initial enhanced image is subjected to targeted re-enhancement. Based on the degradation migration relationship, the region to be re-enhanced and the enhancement direction of the initial enhanced image are determined. The enhancement influence relationship between the region to be re-enhanced and the adjacent effective semantic region is determined by combining spatial correlation. Based on the acceptable enhancement boundary, the allowable adjustment range of local enhancement parameters is determined to obtain the final enhanced image.

2. The image enhancement method based on AI vision as described in claim 1, characterized in that, The steps for acquiring the image to be enhanced, performing unified preprocessing on the image to be enhanced to generate a standardized analysis image, and performing basic statistical analysis on the image to be enhanced to obtain basic statistical information are as follows: The image to be enhanced is obtained, and its size is standardized and its color space is unified to generate a unified image to be enhanced. The unified image to be enhanced is then subjected to brightness extraction and pixel normalization, and preliminary suppression is performed to generate a standardized analysis image. Basic statistical features are extracted from the image to be enhanced through basic statistical analysis, and then normalized and sequentially stitched to generate basic statistical information.

3. The image enhancement method based on AI vision as described in claim 1, characterized in that, The specific steps for performing AI visual analysis on the standardized analysis image and outputting effective semantic region information are as follows: The standardized analysis image is input into the AI ​​visual analysis model, which performs multi-layer visual feature extraction, multi-scale feature mapping and cross-layer fusion processing on the standardized analysis image to generate a semantic analysis feature map. Candidate regions are extracted based on semantic parsing feature maps to generate candidate semantic regions. Then, semantic category discrimination, region boundary localization, and recognition confidence calculation are performed on the candidate semantic regions to select effective semantic regions. Based on the semantic category labels, regional location parameters, and recognition confidence of the effective semantic region, spatial associations and semantic priority relationships are established and encapsulated to form effective semantic region information.

4. The image enhancement method based on AI vision as described in claim 1, characterized in that, The process involves combining basic statistical information to determine the degradation state of regions in the standardized analysis image, identifying the corresponding acceptable enhancement boundaries based on the degradation state, and obtaining a degradation representation token library. The specific steps are as follows: Based on the effective semantic region information, each effective semantic region is located, and the brightness distribution, texture clarity, edge continuity and local contrast changes are extracted. These are then compared with basic statistical information to determine the degradation type and degradation level of each effective semantic region. Based on the semantic category label, degradation type, and degradation level of the effective semantic region, the allowable range of variation corresponding to brightness, contrast, texture, and edge state is determined to obtain the acceptable enhancement boundary; The degradation type, degradation level, and acceptable enhancement boundary of the effective semantic region are associated with the corresponding region location and encoded to generate regional degradation representation tokens. The regional degradation representation tokens are then organized according to the correspondence between effective semantic regions to form a degradation representation token library.

5. The image enhancement method based on AI vision as described in claim 1, characterized in that, The process of generating an enhanced control policy based on valid semantic region information and a degraded representation token library involves the following steps: The effective semantic region information is associated with the degenerate representation token library to output the target enhancement mode and local enhancement parameters; Based on the semantic category labels, degradation types, degradation levels, and acceptable enhancement boundaries of the effective semantic region, the target enhancement mode and local enhancement parameters are configured to generate candidate enhancement configurations. Then, based on spatial correlation and semantic priority, the candidate enhancement configurations are reconciled and ordered to form an enhancement control strategy.

6. The image enhancement method based on AI vision as described in claim 1, characterized in that, The process of using an enhancement control strategy to perform the first image enhancement on the image to be enhanced and outputting a preliminary enhanced image involves the following steps: The image to be enhanced is divided into multiple enhancement regions. According to the target enhancement mode and local enhancement parameters corresponding to each enhancement region, the image to be enhanced is subjected to region adaptive enhancement processing based on the enhancement control strategy to obtain the region enhancement result. The enhancement transition between adjacent regions is coordinated to form transition coordination information, which is then fused to output a preliminary enhanced image.

7. The image enhancement method based on AI vision as described in claim 1, characterized in that, The process involves identifying the original degradation state, residual state, and newly generated degradation state after the first image enhancement, and establishing a degradation transfer relationship based on the changes in degradation state before and after the first image enhancement and the corresponding enhancement operations. The specific steps are as follows: The brightness, texture, edge and local contrast status of each effective semantic region in the preliminary enhanced image are extracted and compared with the corresponding degradation type and degradation level before the first image enhancement. The original degradation fading status and residual status are determined based on the changes in degradation characterization. The enhanced state is compared with the corresponding acceptable enhancement boundary. When the state that did not meet the degradation judgment condition before the first image enhancement exceeds the corresponding acceptable enhancement boundary after enhancement, it is determined to be a newly degraded state, and the corresponding newly degraded type and state change degree are recorded. Based on the changes in degradation state before and after the first image enhancement of each effective semantic region, the corresponding target enhancement mode and local enhancement parameters are associated to determine the correspondence between the original degradation state, residual state and newly generated degradation state after the corresponding enhancement operation, thus forming a degradation migration relationship.

8. The image enhancement method based on AI vision as described in claim 7, characterized in that, The enhancement influence relationship between the region to be enhanced and adjacent effective semantic regions is determined by combining spatial correlation, and the allowable adjustment range of local enhancement parameters is determined according to the acceptable enhancement boundary to obtain the final enhanced image. The specific steps are as follows: The regions corresponding to the residual state and the newly degraded state are identified as regions to be re-enhanced. The original enhancement direction corresponding to the residual state is determined according to the degradation migration relationship, and the reverse suppression direction is determined according to the enhancement operation corresponding to the newly degraded state, thus obtaining the enhancement direction of each region to be re-enhanced. Obtain the remaining change between the current enhancement state of the region to be enhanced and the corresponding acceptable enhancement boundary, and determine the effective semantic region adjacent to the region to be enhanced based on the spatial correlation. Obtain the remaining change, determine the allowable adjustment range of the local enhancement parameters based on the remaining change, and correct the local enhancement parameters that exceed the allowable adjustment range. Directional re-enhancement is performed on each region to be re-enhanced according to the enhancement direction and the corrected local enhancement parameters, and the regions that have been re-enhanced are then fused to obtain the final enhanced image.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the AI ​​vision-based image enhancement method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the AI ​​vision-based image enhancement method according to any one of claims 1 to 8.

Images