An image quality intelligent enhancement method and system based on multi-feature fusion

Abstract

This invention relates to the field of digital image processing technology. It discloses an intelligent image quality enhancement method and system based on multi-feature fusion. The method includes acquiring the image to be enhanced, extracting texture response information, color offset, noise intensity, and edge continuity, constructing a defect representation map to centrally present multiple types of degradation within the same representation, determining feature aggregation areas and scene transition areas based on the defect representation map, generating enhancement constraint values, performing partitioned enhancement and continuous stitching on the image to be enhanced to obtain an intermediate enhancement map, and finally performing residual compensation on the intermediate enhancement map based on the defect representation map to obtain the target enhancement map. This method maintains the integrity of defect representation under strong perturbation degradation conditions and provides stable enhancement results across multiple scenes.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, and more specifically, to an intelligent image quality enhancement method and system based on multi-feature fusion. Background Technology

[0002] Intelligent image defect repair and image quality enhancement technologies are typically applied in image acquisition terminals, image processing platforms, and batch image processing carriers. They are constrained by processing latency, computing resources, and deployment adaptability. Existing technologies usually revolve around local feature determination, global statistical representation, regional correlation measurement, and staged enhancement processing to improve image quality problems such as blurring, noise, color deviation, and detail degradation. These technologies are generally applicable under conditions where the degradation type is relatively simple, the image content changes are limited, and the image quality distribution is relatively stable.

[0003] In real-world image processing scenarios, input images often contain two unstable factors: superimposed degradation interference and changes in scene distribution. These factors can cause deviations in the representation of image defects by existing methods, further affecting the consistency of response during the enhancement process in different scenes. Ultimately, this makes it difficult to maintain stable image quality restoration results. Therefore, the technical problems that need to be solved are how to maintain the integrity of defect representation under strongly perturbed and degraded images, and how to ensure the stability of restoration results under consistency across multiple scenes.

[0004] In view of this, the present invention proposes an intelligent image quality enhancement method and system based on multi-feature fusion to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of the prior art, the present invention provides an intelligent image quality enhancement method and system based on multi-feature fusion.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for intelligent image quality enhancement based on multi-feature fusion is provided, including: The image to be enhanced is obtained, and its texture response information, color offset, noise intensity, and edge continuity are extracted. A defect characterization map is constructed based on the texture response information, color offset, noise intensity, and edge continuity. Based on the defect characterization map, the feature clustering area is determined, and based on the feature clustering area, the scene transition area is determined. Enhanced constraint values ​​are generated based on the defect characterization map, feature clustering area, and scene transition area. The image to be enhanced is partitioned according to the enhancement constraint value to obtain the partition enhancement result. The partition enhancement result is then continuously stitched together according to the scene transition area to obtain the intermediate enhanced image. The intermediate enhancement map is compensated for residuals based on the defect characterization map to obtain the target enhancement map.

[0007] Secondly, an image quality intelligent enhancement system based on multi-feature fusion is provided, which is used to implement the above-mentioned image quality intelligent enhancement method based on multi-feature fusion, including: Flaw construction module: used to acquire the image to be enhanced, extract the texture response information, color offset, noise intensity and edge continuity of the image to be enhanced, and construct a flaw characterization map based on the texture response information, color offset, noise intensity and edge continuity; Region determination module: used to determine feature clustering areas based on the defect characterization map, and to determine scene transition areas based on the feature clustering areas; Constraint generation module: used to generate enhanced constraint values ​​based on the defect characterization map, feature aggregation area, and scene transition area; Partition Enhancement Module: This module is used to perform partition enhancement on the image to be enhanced based on enhancement constraint values, obtain partition enhancement results, and continuously stitch the partition enhancement results according to the scene transition area to obtain an intermediate enhanced image. Residual compensation module: Used to perform residual compensation on the intermediate enhancement map based on the defect characterization map to obtain the target enhancement map.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires the image to be enhanced, extracts texture response information, color offset, noise intensity, and edge continuity, and constructs a defect characterization map based on these features, so that blurring, noise, color deviation, and edge attenuation are presented in the same characterization map. Then, feature clustering areas and scene transition areas are determined based on the defect characterization map, thus distinguishing and constraining the degradation concentration areas and scene transition areas. Further, enhancement constraint values ​​are generated based on the defect characterization map, feature clustering areas, and scene transition areas, ensuring a unified basis for the enhancement intensity of different areas. Subsequently, the image to be enhanced is partitioned according to the enhancement constraint values ​​to obtain partitioned enhancement results. The partitioned enhancement results are then continuously stitched together based on the scene transition areas to obtain an intermediate enhancement map, reducing enhancement jumps between different scenes. Finally, residual compensation is performed on the intermediate enhancement map based on the defect characterization map to obtain the target enhancement map. This maintains the integrity of defect characterization under strong perturbation degradation conditions and stabilizes the enhancement results under multi-scene consistency constraints. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an intelligent image quality enhancement method based on multi-feature fusion according to the present invention. Figure 2 This is a schematic diagram of the structure of an image quality intelligent enhancement system based on multi-feature fusion according to the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. In the following detailed description, many specific details are set forth to provide a thorough understanding of the exemplary embodiments described. However, it will be apparent to those skilled in the art that the described embodiments may be practiced without some or all of these specific details. In other exemplary embodiments, well-known structures have not been described in detail to avoid unnecessarily obscuring the concepts of this disclosure. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention. Furthermore, the various aspects described in the embodiments may be combined arbitrarily without conflict.

[0011] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0012] Example 1 Figure 1 This disclosure illustrates at least one embodiment of an image quality intelligent enhancement method based on multi-feature fusion, comprising: S10: Obtain the image to be enhanced, extract the texture response information, color offset, noise intensity and edge continuity of the image to be enhanced, and construct a defect characterization map based on the texture response information, color offset, noise intensity and edge continuity. In this embodiment, texture response information is used to characterize the detail fluctuation state of a local area in the image to be enhanced, color offset is used to characterize the deviation state of the local area relative to the reference color distribution, noise intensity is used to characterize the degree of random interference in the local area, edge continuity is used to characterize the integrity of contour connection in the local area, and the defect characterization map is the distribution result formed by arranging the above information according to the positional relationship of the local sampling blocks in the image to be enhanced. Those skilled in the art can identify the image quality anomaly type and its spatial distribution corresponding to each local sampling block based on the defect characterization map.

[0013] It should be noted that the process of first extracting texture response information, color offset, noise intensity, and edge continuity, and then constructing a defect characterization map, focuses on incorporating structural degradation information and interference degradation information in local areas into the same characterization system. In strongly perturbed degraded images, blurring, color shift, noise, and edge breaks often appear intertwined. If a single feature is used for judgment, it is easy to misjudge real details as interference information, and vice versa. However, by uniformly arranging the above information through the defect characterization map, a common input basis can be provided for subsequent feature cluster area recognition.

[0014] For example, the image to be enhanced could be a promotional image of a game character. In this image, the character's hair area has weakened edge continuity, the clothing pattern area has weakened texture response information, the background lighting effect area has increased color shift, and the dark shadow area has increased noise intensity. The corresponding local sampling blocks can be written as b1, b2, b3, and b4, where b1 corresponds to the character's hair area, b2 corresponds to the clothing pattern area, b3 corresponds to the background lighting effect area, and b4 corresponds to the dark shadow area. The defect characterization map formed by writing the texture response information, color shift, noise intensity, and edge continuity corresponding to each local sampling block according to their original positions represents the distribution of local defects in the image to be enhanced.

[0015] A defect characterization map is constructed based on texture response information, color offset, noise intensity, and edge continuity, including: The image to be enhanced is divided into multiple local sampling blocks, and the texture response information, color offset, noise intensity and edge continuity corresponding to each local sampling block are extracted, as well as the positional relationship of each local sampling block in the image to be enhanced. The texture response information and edge continuity are combined to form a structural identifier, and the color offset and noise intensity are combined to form an interference identifier. Based on the structural identifier and interference identifier, the defect signature identifier corresponding to each local sampling block is determined. The defect signature identifiers are arranged according to the positional relationship of each local sampling block to obtain the defect characterization map.

[0016] In this embodiment, a local sampling block can be understood as a basic analysis unit obtained after spatially dividing the image to be enhanced. Its function is to decompose the complex image quality problems in the entire image to be enhanced into multiple local regions for separate representation. The positional relationship is used to record the original distribution position of each local sampling block in the image to be enhanced, thereby ensuring that the defect signature identifier generated later can be written back to the corresponding region. Those skilled in the art can understand that extracting texture response information, color offset, noise intensity and edge continuity, while retaining the positional relationship of each local sampling block, aims to retain the spatial distribution basis while extracting local features, so as to construct a defect representation map with spatial orientation in the future.

[0017] It should be noted that combining texture response information with edge continuity to form a structural identifier is to combine the representation of the detail undulation state and contour connection state in the local sampling block. Combining color offset and noise intensity to form an interference identifier is to combine the representation of color anomalies and random interference in the local sampling block. This processing method is not a simple superposition of features, but rather classifies structural degradation and interference degradation into two categories of representation results. This allows for the subsequent determination of defect categories for each local sampling block, distinguishing both the presence of detail attenuation and edge breakage in the local area, as well as the presence of color shift and noise accumulation in the local area.

[0018] Understandably, structural identifiers are used to characterize the abnormal state of local sampling blocks at the structural level, interference identifiers are used to characterize the abnormal state of local sampling blocks at the interference level, and defect signature identifiers are local defect category markers formed based on structural identifiers and interference identifiers. The defect signature identifiers are arranged according to the positional relationship of each local sampling block. Essentially, this remaps the local defect category back to its original spatial location in the image to be enhanced, so that the defect representation map not only reflects what type of image quality abnormality exists in each local sampling block, but also reflects the distribution location of this type of image quality abnormality in the image to be enhanced, thus providing a basis for subsequent identification of feature clusters.

[0019] For example, using the aforementioned example of a game character promotional image, the image to be enhanced is divided into local sampling blocks b1, b2, b3, and b4. Here, b1 corresponds to the character's hair area, b2 corresponds to the clothing pattern area, b3 corresponds to the background lighting effect area, and b4 corresponds to the dark shadow area. If the texture response information of b1 is weak and the edge continuity is weak, then the structural identifier corresponding to b1 is characterized as an edge attenuation type. If the color shift of b3 is strong and the noise intensity of b4 is strong, then the interference identifiers corresponding to b3 and b4 are characterized as color shift type and noise interference type, respectively. Based on this, the defect signature identifiers corresponding to b1, b2, b3, and b4 are formed, and after arranging them according to their original positions, the defect characterization map of the image to be enhanced can be obtained.

[0020] Based on the structural identifier and interference identifier, the defect signature identifier corresponding to each local sampling block is determined, including: Structural identifiers are classified into texture-dominant, edge-dominant, and texture-and-edge coexistence types, while interference identifiers are classified into color-bias-dominant, noise-dominant, and color-and-noise coexistence types. Retrieve the structure identifier category and interference identifier category corresponding to each local sampling block, and select the corresponding signature category according to the pairing relationship between the structure identifier category and the interference identifier category; Assign the signature category to the corresponding local sampling block to obtain the defective signature identifier for each local sampling block.

[0021] In this embodiment, texture-dominant type can be understood as the category of structural identifiers whose main representation is texture undulation; edge-dominant type can be understood as the category of structural identifiers whose main representation is contour connection change; texture and edge coexistence type can be understood as the category where texture undulation change and contour connection change coexist; color shift-dominant type can be understood as the category of interference identifiers whose main representation is color shift; noise-dominant type can be understood as the category of interference identifiers whose main representation is random interference; color and noise coexistence type can be understood as the category where color shift and random interference coexist; and defect signature identifier is a local defect category mark obtained by mapping the structural identifier category to the interference identifier category.

[0022] It should be noted that the purpose of classifying structural identifiers into texture-dominant, edge-dominant, and texture-and-edge coexistence types, and the purpose of classifying interference identifiers into color-biased, noise-dominant, and color-noise coexistence types, is to first classify structural anomalies and interference anomalies in local sampling blocks into limited categories, and then determine the defect signature identifier through the pairing relationship between categories. After this processing, the local sampling block can not only reflect whether there is structural degradation or interference degradation, but also reflect the combination relationship between the two, thereby avoiding mixing image quality anomalies from different sources into a single judgment result.

[0023] The corresponding signature category is selected based on the pairing relationship between the structural identifier category and the interference identifier category. The key point of the processing is to compress the structural layer anomalies and interference layer anomalies in the local sampling block into a single defect signature identifier. Based on this, those skilled in the art can directly identify the composite defect type of the local sampling block based on the defect signature identifier, without having to repeatedly retrieve texture response information, color offset, noise intensity and edge continuity in subsequent processing. This makes the defect characterization map more readable and more directly directional.

[0024] For example, using the aforementioned example of game character promotional images, if local sampling block b1 corresponds to the character's hair area, and its structural identifier category is edge-dominant and its interference identifier category is noise-dominant, then the signature category corresponding to b1 is characterized as an edge-noise composite category. If local sampling block b2 corresponds to the clothing pattern area, and its structural identifier category is texture-dominant and its interference identifier category is color-bias-dominant, then the signature category corresponding to b2 is characterized as a texture-color-bias composite category. If local sampling block b3 corresponds to the background lighting effect area, and its structural identifier category is texture-edge coexistence and its interference identifier category is color-noise coexistence, then the signature category corresponding to b3 is characterized as a texture-edge-color-noise composite category. After assigning each signature category to the corresponding local sampling block, the defect signature identifier corresponding to each local sampling block can be obtained.

[0025] Furthermore, compared to the processing method that directly judges the image quality of local sampling blocks based on single features, the above processing method first forms structural identifier categories and interference identifier categories, and then generates defect signature identifiers through pairing relationships. This allows composite defects in local sampling blocks to be uniformly described in a categorized manner. In strongly perturbed degraded images, this processing method helps reduce the confusion between different defect types and facilitates subsequent continuous processing around the generation of feature aggregation areas, scene transition areas, and enhancement constraint values.

[0026] S20: Determine the feature clustering area based on the defect characterization map, and determine the scene transition area based on the feature clustering area; Based on the defect characterization map, feature clustering regions are identified, including: Local sampling blocks with the same structural identifier are extracted from the defect characterization map to form a structural association set, and local sampling blocks with the same interference identifier are extracted to form an interference association set. The structural correlation set and the interference correlation set are spatially superimposed to form a cluster candidate band, which defines the enclosure range of the local sampling block; Connect the local sampling blocks enclosed by the clustered candidate bands to obtain the feature clustering region.

[0027] In this embodiment, the feature clustering area can be understood as the region formed by the continuous distribution of similar defect types in the defect representation map, and the scene transition area can be understood as the region located between different feature clustering areas that affects the enhancement connection. The process of determining the feature clustering area based on the defect representation map and then determining the scene transition area based on the feature clustering area is to first identify the concentrated distribution location of the composite defect and then identify the connection range between adjacent areas. This allows the subsequent enhancement processing to not only focus on the main defect area but also take into account the transition relationship between different areas. Local sampling blocks with the same structural identifier are extracted from the defect representation map to form a structural association set, and local sampling blocks with the same interference identifier are extracted to form an interference association set. The purpose is not to process structural anomalies and interference anomalies separately, but to first find local sampling blocks from the two types of representation results that may spatially point to the same type of defect area. The former reflects the continuous anomalies of the local area at the texture and edge level, and the latter reflects the continuous anomalies of the local area at the color and noise level. Both serve as the basis for subsequent identification of the clustering range.

[0028] It is understandable that spatially superimposing the structural association set and the interference association set to form a cluster candidate band essentially uses the spatial overlap of the two types of association sets to limit the outer range of the concentrated distribution of defects. The cluster candidate band does not directly correspond to the feature cluster area. Instead, the enclosing boundary is given first, and then the feature cluster area is determined by connecting the local sampling blocks enclosed by it. This approach helps to avoid local gaps or regional breaks when dividing the area based on a single identifier, making the obtained feature cluster area more consistent with the actual distribution pattern of composite defects.

[0029] For example, using the aforementioned example of a game character promotional image, local sampling block b1 corresponds to the character's hair area, local sampling block b2 corresponds to the clothing pattern area, local sampling block b3 corresponds to the background lighting effect area, and local sampling block b4 corresponds to the dark shadow area. If b1 and b2 have similar structural identifiers, and b3 and b4 have similar interference identifiers, then b1 and b2 can form a structural association set, and b3 and b4 can form an interference association set. When the structural association set and the interference association set form an enclosing relationship at a local position in the image, a clustering candidate band can be formed accordingly, and the local sampling blocks enclosed by the clustering candidate band can be further connected to obtain the corresponding feature clustering area.

[0030] Spatially overlaying the structural association set and the interference association set forms a clustering candidate band, including: In the structural association set, extract the local sampling blocks that are in contact with the outer local sampling blocks to form the structural boundary set, and in the interference association set, extract the local sampling blocks that are in contact with the outer local sampling blocks to form the interference boundary set. The structural boundary set and the interference boundary set are aligned in position, and local sampling blocks that fall within the boundary range of both are extracted to form a shared edge distribution region. By connecting the shared-edge distribution areas according to their contact relationships, we can obtain the candidate aggregation zone.

[0031] In this embodiment, the structural boundary set can be understood as the set of local sampling blocks in the structural association set that are in direct contact with the non-structural association region, the interference boundary set can be understood as the set of local sampling blocks in the interference association set that are in direct contact with the non-interference association region, the shared edge distribution region can be understood as a local region that is simultaneously defined by the structural boundary set and the interference boundary set, and the aggregation candidate band is a strip-shaped enclosed region formed by connecting the shared edge distribution regions along the contact direction. Those skilled in the art can understand that this process does not directly overlap the structural association set and the interference association set as a whole, but first extracts the boundaries of the two sets, and then uses the positional correspondence between the boundaries to determine the outer edge range that is more suitable for enclosing the composite defect region.

[0032] It should be noted that the purpose of extracting local sampling blocks that contact the outer local sampling blocks from the structural association set to form the structural boundary set, and extracting local sampling blocks that contact the outer local sampling blocks from the interference association set to form the interference boundary set, is to first separate the local sampling blocks that actually participate in the definition of the region boundary from the two types of association sets. This is because the outer contour of the feature clustering area is mainly determined by the boundary position. If the structural association set and the interference association set are directly superimposed as a whole, it is easy to include the local sampling blocks located inside in the boundary judgment, thereby weakening the limiting effect of the clustering candidate band on the enclosure range.

[0033] Understandably, by aligning the structural boundary set with the interference boundary set and extracting local sampling blocks that fall within the boundaries of both sets to form a shared-edge distribution region, the key is to find the areas that the structural anomaly boundary and the interference anomaly boundary pass through in space. Then, the shared-edge distribution regions are connected in series according to their contact relationship to obtain a cluster candidate band. The cluster candidate band obtained in this way retains the boundary trends of both the structural association set and the interference association set, and is therefore more suitable for characterizing the peripheral distribution pattern of the composite defect region.

[0034] For example, using the aforementioned example of a game character promotional image, local sampling block b1 corresponds to the character's hair area, local sampling block b2 corresponds to the clothing pattern area, local sampling block b3 corresponds to the background lighting effect area, and local sampling block b4 corresponds to the dark shadow area. If the structural association set formed by b1 and b2 extracts the structural boundary set at the outer edge, and the interference association set formed by b3 and b4 extracts the interference boundary set at the outer edge, and both types of boundaries pass through the boundary area between the character outline and the background at the local position of the image, then the local sampling blocks corresponding to this boundary area can form a common edge distribution area. After connecting this common edge distribution area according to the contact relationship, the aggregation candidate band can be obtained.

[0035] Furthermore, compared to the method of directly spatially overlapping the entire associated region, the above method first extracts the structural boundary set and the interference boundary set, then forms a common edge distribution region through positional alignment, and further obtains the cluster candidate band. This allows the cluster candidate band to better reflect the common distribution position of the structural abnormal boundary and the interference abnormal boundary. In strongly perturbed degraded images, this method helps to reduce the interference of internal non-boundary local sampling blocks on the region enclosure judgment, and helps to make the subsequently obtained feature cluster area closer to the actual outer edge of the composite defect region.

[0036] S30: Generate enhanced constraint values ​​based on the defect characterization map, feature clustering area, and scene transition area; Enhanced constraint values ​​are generated based on the defect characterization map, feature clustering areas, and scene transition areas, including: Extract the defect signature identifiers that fall into the feature cluster area from the defect characterization map, and extract the defect signature identifiers that come into contact with the scene transition area to form a region signature set; Based on the distribution relationship between the regional signature set and the feature clustering area, the clustering constraint component is determined, and based on the adjacency relationship between the regional signature set and the scene transition area, the transition constraint component is determined. Generate a set of regional constraint indexes according to the correspondence between cluster constraint components and transition constraint components; Map the set of region constraint indexes to enhanced constraint values ​​to obtain the enhanced constraint values.

[0037] In this embodiment, the enhancement constraint value can be understood as the enhancement control basis set for different local regions in the image to be enhanced. It does not directly correspond to a single feature, but is jointly defined by the defect characterization map, the feature aggregation area, and the scene transition area. The defect characterization map reflects the distribution of defect categories in the local sampling block, the feature aggregation area reflects the concentrated location of compound defects, and the scene transition area reflects the connection range between different regions. Therefore, generating enhancement constraint values ​​based on the defect characterization map, the feature aggregation area, and the scene transition area is essentially to first determine the main enhancement area and then determine the enhancement connection relationship between the main enhancement area and the surrounding transition area.

[0038] It should be noted that the defect signature identifiers falling into the feature clustering area are extracted from the defect characterization map, and the defect signature identifiers that come into contact with the scene transition area are extracted to form a regional signature set. The purpose is to include the local defect categories in the main defect area and the local defect categories in the transition area within the same analysis scope. Then, the clustering constraint component is determined according to the distribution relationship between the regional signature set and the feature clustering area, and the transition constraint component is determined according to the adjacency relationship between the regional signature set and the scene transition area. The regional constraint index set is generated according to the correspondence between the clustering constraint component and the transition constraint component, and then the enhancement constraint value is mapped to obtain the enhancement constraint value. Thus, the enhancement constraint value reflects both the enhancement requirements of the main defect area and the connection requirements at the regional boundary.

[0039] For example, using the aforementioned game character promotional image, local sampling block b1 corresponds to the character's hair area, local sampling block b2 corresponds to the clothing pattern area, local sampling block b3 corresponds to the background lighting effect area, and local sampling block b4 corresponds to the dark shadow area. If b1 and b2 fall into the feature clustering area, b3 is in contact with the scene transition area, and b4 is located outside the feature clustering area, then the defect signature identifiers corresponding to b1 and b2 and the defect signature identifier corresponding to b3 can be extracted from the defect characterization image to form a region signature set. Then, based on the distribution relationship of b1 and b2 in the feature clustering area, clustering constraint components are formed, and based on the adjacency relationship between b3 and the scene transition area, transition constraint components are formed. Furthermore, a region constraint index set and enhanced constraint values ​​are generated.

[0040] A set of region constraint indexes is generated based on the correspondence between cluster constraint components and transition constraint components, including: Based on the defect signature identifiers corresponding to each local sampling block within the feature clustering region, extract the clustering dominant signature; Extract the transition-related signature based on the defect signature identifiers corresponding to each local sampling block within the scene transition area; Generate a set of region constraint indexes based on the association between the cluster-dominant signature and the transitional accompanying signature.

[0041] In this embodiment, the cluster dominant signature can be understood as a defective signature identifier that plays a major limiting role in the direction of regional enhancement within the feature cluster area, and the transition accompanying signature can be understood as a defective signature identifier that plays a cooperating role in the direction of regional connection within the scene transition area. The regional constraint index set is the index result formed by associating the cluster dominant signature and the transition accompanying signature according to a preset correspondence method. The above processing does not process the feature cluster area and the scene transition area separately in isolation, but first extracts the dominant enhancement basis from the feature cluster area, and then extracts the connection enhancement basis from the scene transition area, so that the regional constraint index set has both the internal enhancement direction of the region and the transition direction of the region boundary.

[0042] For example, using the aforementioned example of game character promotional images, local sampling blocks b1 and b2 fall into the feature aggregation area, and local sampling block b3 is located in the scene transition area. If the defect signatures corresponding to b1 and b2 mainly represent the attenuation of hair edges and the weakening of clothing patterns, then the aggregation-dominant signature can be extracted from b1 and b2. If the defect signatures corresponding to b3 represent the background light effect shift, then the transition-accompanying signature can be extracted from b3. Then, according to the association relationship between the aggregation-dominant signature and the transition-accompanying signature, a set of regional constraint indexes is generated, so that the subsequently obtained enhanced constraint values ​​can reflect both the main repair needs of the main character area and the transition connection needs between the main character area and the background light effect area.

[0043] Based on the defect signature identifiers corresponding to each local sampling block within the scene transition area, extract the transition-accompanying signature, including: Extract and arrange the defect signature identifiers corresponding to the local sampling blocks along the extension direction of the scene transition area; Retrieve the signature categories that are adjacent to the dominant signature in the defective signature identifier arrangement to form a candidate accompanying signature set; Extract transitional signatures based on the continuous distribution of candidate accompanying signature sets in the scene transition zone.

[0044] In this embodiment, the transition-accompanying signature can be understood as a defect signature identifier that plays a coordinating role in the direction of regional connection within the scene transition area. The key to its extraction is not simply selecting any signature category in the scene transition area, but observing the arrangement of defect signature identifiers along the extension direction of the scene transition area. This is because the scene transition area itself corresponds to the connection range between different areas. Extracting the arrangement of defect signature identifiers corresponding to local sampling blocks along its extension direction can reflect the spatial transmission relationship of defect categories within the scene transition area, thereby providing a basis for subsequent identification of the signature categories that truly participate in regional connection.

[0045] It should be noted that the purpose of retrieving the signature categories that connect with the dominant cluster signature in the defective signature identifier arrangement to form a candidate accompanying signature set is to screen out defective signature identifiers that have a direct connection relationship with the feature cluster area from the scene transition area. This is because multiple local defect categories may exist simultaneously in the scene transition area. If the cluster dominant signature is not used for limitation, it is easy to include signature categories that have no direct connection relationship with the main enhancement area. However, by extracting transition accompanying signatures according to the continuous distribution relationship of the candidate accompanying signature set in the scene transition area, the obtained transition accompanying signatures can more accurately reflect the connection category corresponding to the outward transition of the main enhancement area.

[0046] For example, using the aforementioned game character promotional image as an example, local sampling blocks b1 and b2 fall into the feature aggregation region, while local sampling block b3 is located in the scene transition region. Their correspondence can be seen in Table 1: Table 1. Relationship between local sampling blocks and transitional signature extraction Furthermore, according to Table 1, the signature categories corresponding to b1 and b2 serve as the extraction source of the cluster-dominant signature, and the signature category corresponding to b3 is connected to the cluster-dominant signature and is continuously distributed in the scene transition area. Therefore, the signature category corresponding to b3 is extracted as the transition accompanying signature, so that the subsequently generated regional constraint index set can not only enhance and control the main flaws of the main character area, but also take into account the transition between the main character area and the background lighting effect area.

[0047] S40: Perform partition enhancement on the image to be enhanced according to the enhancement constraint value to obtain the partition enhancement result, and continuously stitch the partition enhancement result according to the scene transition area to obtain the intermediate enhancement image; In this embodiment, the image to be enhanced is partitioned according to the enhancement constraint value. This can be understood as performing corresponding enhancement processing on different regions of the image to be enhanced based on the enhancement control criteria corresponding to different local regions, and obtaining the partitioned enhancement result. The enhancement constraint value does not require each region to use the same enhancement method, but is used to limit the enhancement focus of each region, so that the local sampling blocks in the feature aggregation area prioritize the repair of the main defects, and the local sampling blocks near the scene transition area are adjusted to take into account the regional connection relationship. Thus, the partitioned enhancement result reflects the repair needs of the main defect area and also retains the transition basis between adjacent areas.

[0048] It should be noted that the intermediate enhanced image is obtained by continuously stitching the enhancement results of each region according to the scene transition area. The key point of this process is that it does not directly write back the enhancement results of each region side by side. Instead, it connects the boundary connection relationship of the enhancement results of adjacent regions according to the spatial range of the scene transition area in the image to be enhanced. This is because if the enhancement results of each region are output independently, it is easy to form brightness jumps, color breaks or contour misalignments at the region boundaries. However, by continuously stitching through the scene transition area, the enhancement results of adjacent regions can remain coherent within the transition range, thus obtaining an intermediate enhanced image that can be used for subsequent residual compensation processing.

[0049] For example, using the aforementioned example of a game character promotional image, local sampling blocks b1 and b2 fall into the feature aggregation area, local sampling block b3 is located in the scene transition area, and local sampling block b4 is located in the outer area. Based on the enhancement constraint value, hair edge repair and clothing pattern enhancement can be performed preferentially on the corresponding areas of b1 and b2. The light and shadow transition connecting the corresponding area of ​​b3 with the main area can be preserved, and the overall coordination of the dark area of ​​the background can be maintained on the corresponding area of ​​b4. Then, based on the scene transition area, the partition enhancement results between b2 and b3 and between b3 and b4 are continuously stitched together to obtain an intermediate enhancement image. In this way, the main flaw repair of the main character area and the transition of the background lighting area can be kept coordinated in the same image result.

[0050] S50: Perform residual compensation on the intermediate enhancement map based on the defect characterization map to obtain the target enhancement map.

[0051] Based on the defect characterization map, residual compensation is performed on the intermediate enhancement map to obtain the target enhancement map, including: Extract the structural and interference identifiers corresponding to each local sampling block from the defect characterization map, and extract the enhanced image fragments of the corresponding local sampling blocks from the intermediate enhancement map; Structural residual fragments are generated from the enhanced image fragments based on structural identifiers, and interference residual fragments are generated from the enhanced image fragments based on interference identifiers. The structural residual fragments and interference residual fragments are written back to the corresponding local sampling blocks to obtain the target enhancement map.

[0052] In this embodiment, residual compensation is performed on the intermediate enhancement image based on the defect characterization image. This can be understood as, after the partition enhancement and continuous stitching are completed, the structural and interference identifiers retained in the defect characterization image are used to perform targeted adjustments on local areas in the intermediate enhancement image. The structural identifiers correspond to the details and contours of the local area, and the interference identifiers correspond to the color and noise of the local area. Therefore, the structural and interference identifiers corresponding to each local sampling block are extracted from the defect characterization image, and the enhanced image fragments corresponding to the local sampling blocks are extracted from the intermediate enhancement image. The purpose is to establish the residual compensation on a one-to-one correspondence between the local defect category and the local enhancement result, thereby avoiding indiscriminate adjustments to the intermediate enhancement image.

[0053] It should be noted that generating structural residual fragments from the enhanced image segments based on structural identifiers and generating interference residual fragments from the enhanced image segments based on interference identifiers focuses on extracting the structural and interference biases that remain in the intermediate enhanced image. Then, the structural and interference residual fragments are written back to the corresponding local sampling blocks to obtain the target enhanced image. After this processing, issues such as contour breaks and texture reduction corresponding to structural identifiers can be compensated for in local areas, and issues such as color shifts and noise residues corresponding to interference identifiers can also be corrected in local areas. This makes the target enhanced image more convergent in terms of local details and regional coordination compared to the intermediate enhanced image.

[0054] For example, using the aforementioned game character promotional image as an example, local sampling block b1 corresponds to the character's hair area, local sampling block b2 corresponds to the clothing pattern area, local sampling block b3 corresponds to the background lighting effect area, and local sampling block b4 corresponds to the dark shadow area. In the intermediate enhancement image, if b1 still has insufficient local edge connection, b2 still has a lack of pattern details, b3 still has a color cast in the lighting effect, and b4 still has residual interference in the dark area, then structural residual fragments can be generated based on the structural identifiers corresponding to b1 and b2, and interference residual fragments can be generated based on the interference identifiers corresponding to b3 and b4. Then, the corresponding residual fragments are written back to the local sampling blocks where b1, b2, b3, and b4 are located to obtain the target enhancement image. In this way, the detail repair of the main character area and the color and noise correction of the background area are kept coordinated in the same output result.

[0055] Furthermore, compared to the processing method that directly outputs the result after partition enhancement, the above processing method, based on the intermediate enhanced image, combines the structural and interference markers in the defect characterization image for residual compensation. This allows the local structural deviations and local interference deviations that were not fully repaired during the partition enhancement process to be re-identified and written back in a directional manner. In strongly perturbed degraded images, this processing method helps to reduce the residual contour discontinuities, insufficient detail compensation, residual color shifts, and residual dark interference after region stitching, thereby enabling the target enhanced image to maintain a more stable output effect in terms of local detail restoration and overall image coordination.

[0056] Example 2 Please see Figure 2 As shown, based on the same inventive concept, this embodiment discloses an image quality intelligent enhancement system based on multi-feature fusion. For details not covered in this embodiment, please refer to the relevant parts of Embodiment 1. The system includes: Flaw construction module: used to acquire the image to be enhanced, extract the texture response information, color offset, noise intensity and edge continuity of the image to be enhanced, and construct a flaw characterization map based on the texture response information, color offset, noise intensity and edge continuity; Region determination module: used to determine feature clustering areas based on the defect characterization map, and to determine scene transition areas based on the feature clustering areas; Constraint generation module: used to generate enhanced constraint values ​​based on the defect characterization map, feature aggregation area, and scene transition area; Partition Enhancement Module: This module is used to perform partition enhancement on the image to be enhanced based on enhancement constraint values, obtain partition enhancement results, and continuously stitch the partition enhancement results according to the scene transition area to obtain an intermediate enhanced image. Residual compensation module: Used to perform residual compensation on the intermediate enhancement map based on the defect characterization map to obtain the target enhancement map.

[0057] The detailed description above, in conjunction with the accompanying drawings, describes examples but does not represent all examples that can be implemented or fall within the scope of the claims. The terms “example” and “exemplary” are used in this specification to mean “serving as an example, instance or illustration” and do not mean “superior to or better than other examples”.

[0058] Throughout this specification, the phrase "an embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. Therefore, the use of these phrases may refer to more than one embodiment. Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0059] It should also be noted that these embodiments may be described as processes depicted as flowcharts, structural diagrams, or block diagrams. Although a flowchart may describe the operations as sequential processes, many of these operations can be performed in parallel or concurrently, and the order of these operations may be rearranged.

Claims

1. An intelligent image quality enhancement method based on multi-feature fusion, characterized in that, include: The image to be enhanced is obtained, and its texture response information, color offset, noise intensity, and edge continuity are extracted. A defect characterization map is constructed based on the texture response information, color offset, noise intensity, and edge continuity. Based on the defect characterization map, the feature clustering area is determined, and based on the feature clustering area, the scene transition area is determined. Enhanced constraint values ​​are generated based on the defect characterization map, feature clustering area, and scene transition area. The image to be enhanced is partitioned according to the enhancement constraint value to obtain the partition enhancement result. The partition enhancement result is then continuously stitched together according to the scene transition area to obtain the intermediate enhanced image. The intermediate enhancement map is compensated for residuals based on the defect characterization map to obtain the target enhancement map.

2. The image quality intelligent enhancement method based on multi-feature fusion according to claim 1, characterized in that, A defect characterization map is constructed based on texture response information, color offset, noise intensity, and edge continuity, including: The image to be enhanced is divided into multiple local sampling blocks, and the texture response information, color offset, noise intensity and edge continuity corresponding to each local sampling block are extracted, as well as the positional relationship of each local sampling block in the image to be enhanced. The texture response information and edge continuity are combined to form a structural identifier, and the color offset and noise intensity are combined to form an interference identifier. Based on the structural identifier and interference identifier, the defect signature identifier corresponding to each local sampling block is determined. The defect signature identifiers are arranged according to the positional relationship of each local sampling block to obtain the defect characterization map.

3. The image quality intelligent enhancement method based on multi-feature fusion according to claim 2, characterized in that, Based on the structural identifier and interference identifier, the defect signature identifier corresponding to each local sampling block is determined, including: Structural identifiers are classified into texture-dominant, edge-dominant, and texture-and-edge coexistence types, while interference identifiers are classified into color-bias-dominant, noise-dominant, and color-and-noise coexistence types. Retrieve the structure identifier category and interference identifier category corresponding to each local sampling block, and select the corresponding signature category according to the pairing relationship between the structure identifier category and the interference identifier category; Assign the signature category to the corresponding local sampling block to obtain the defective signature identifier for each local sampling block.

4. The image quality intelligent enhancement method based on multi-feature fusion according to claim 2, characterized in that, Based on the defect characterization map, feature clustering regions are identified, including: Local sampling blocks with the same structural identifier are extracted from the defect characterization map to form a structural association set, and local sampling blocks with the same interference identifier are extracted to form an interference association set. The structural correlation set and the interference correlation set are spatially superimposed to form a cluster candidate band, which defines the enclosure range of the local sampling block; Connect the local sampling blocks enclosed by the clustered candidate bands to obtain the feature clustering region.

5. The image quality intelligent enhancement method based on multi-feature fusion according to claim 4, characterized in that, Spatially overlaying the structural association set and the interference association set forms a clustering candidate band, including: In the structural association set, extract the local sampling blocks that are in contact with the outer local sampling blocks to form the structural boundary set, and in the interference association set, extract the local sampling blocks that are in contact with the outer local sampling blocks to form the interference boundary set. The structural boundary set and the interference boundary set are aligned in position, and local sampling blocks that fall within the boundary range of both are extracted to form a shared edge distribution region. By connecting the shared-edge distribution areas according to their contact relationships, we can obtain the candidate aggregation zone.

6. The image quality intelligent enhancement method based on multi-feature fusion according to claim 4, characterized in that, Enhanced constraint values ​​are generated based on the defect characterization map, feature clustering areas, and scene transition areas, including: Extract the defect signature identifiers that fall into the feature cluster area from the defect characterization map, and extract the defect signature identifiers that come into contact with the scene transition area to form a region signature set; Based on the distribution relationship between the regional signature set and the feature clustering area, the clustering constraint component is determined, and based on the adjacency relationship between the regional signature set and the scene transition area, the transition constraint component is determined. Generate a set of regional constraint indexes according to the correspondence between cluster constraint components and transition constraint components; Map the set of region constraint indexes to enhanced constraint values ​​to obtain the enhanced constraint values.

7. The image quality intelligent enhancement method based on multi-feature fusion according to claim 6, characterized in that, A set of region constraint indexes is generated based on the correspondence between cluster constraint components and transition constraint components, including: Based on the defect signature identifiers corresponding to each local sampling block within the feature clustering region, extract the clustering dominant signature; Extract the transition-related signature based on the defect signature identifiers corresponding to each local sampling block within the scene transition area; Generate a set of region constraint indexes based on the association between the cluster-dominant signature and the transitional accompanying signature.

8. The image quality intelligent enhancement method based on multi-feature fusion according to claim 7, characterized in that, Based on the defect signature identifiers corresponding to each local sampling block within the scene transition area, extract the transition-accompanying signature, including: Extract and arrange the defect signature identifiers corresponding to the local sampling blocks along the extension direction of the scene transition area; Retrieve the signature categories that are adjacent to the dominant signature in the defective signature identifier arrangement to form a candidate accompanying signature set; Extract transitional signatures based on the continuous distribution of candidate accompanying signature sets in the scene transition zone.

9. The image quality intelligent enhancement method based on multi-feature fusion according to claim 2, characterized in that, Based on the defect characterization map, residual compensation is performed on the intermediate enhancement map to obtain the target enhancement map; including: Extract the structural and interference identifiers corresponding to each local sampling block from the defect characterization map, and extract the enhanced image fragments of the corresponding local sampling blocks from the intermediate enhancement map; Structural residual fragments are generated from the enhanced image fragments based on structural identifiers, and interference residual fragments are generated from the enhanced image fragments based on interference identifiers. The structural residual fragments and interference residual fragments are written back to the corresponding local sampling blocks to obtain the target enhancement map.

10. An image quality intelligent enhancement system based on multi-feature fusion, used to implement the image quality intelligent enhancement method based on multi-feature fusion as described in any one of claims 1-9, characterized in that, include: Flaw construction module: used to acquire the image to be enhanced, extract the texture response information, color offset, noise intensity and edge continuity of the image to be enhanced, and construct a flaw characterization map based on the texture response information, color offset, noise intensity and edge continuity; Region Determination Module: Used to determine feature clustering areas based on the defect characterization map; Determine the scene transition zone based on the feature clustering area; Constraint generation module: used to generate enhanced constraint values ​​based on the defect characterization map, feature aggregation area, and scene transition area; Partition Enhancement Module: Used to perform partition enhancement on the image to be enhanced based on enhancement constraint values, and obtain the partition enhancement result; The enhancement results of the partitions are then continuously stitched together based on the scene transition area to obtain an intermediate enhancement image; Residual compensation module: Used to perform residual compensation on the intermediate enhancement map based on the defect characterization map to obtain the target enhancement map.

Images