An image enhancement method, computer device, and storage medium based on information divide-and-conquer

Abstract

This invention discloses an image enhancement method, computer device, and storage medium based on information divide-and-conquer. The invention includes the following steps: decomposing the image to be processed into multiple image blocks; classifying each image block using an image block classification algorithm based on the gray-level co-occurrence matrix entropy to obtain category labels; enhancing each image block according to its category labels based on an early termination mechanism to obtain corresponding enhancement feature information for each image block; and aggregating the enhancement feature information based on feature skip connections to obtain the enhanced result image. This invention employs an information divide-and-conquer strategy, using an image block classification algorithm based on the gray-level co-occurrence matrix entropy to classify image blocks into multiple categories based on their information content. Guided by the early termination mechanism, it uses network structures of varying complexity to enhance image blocks with different information contents, achieving good performance while maintaining low computational cost. This invention is applicable to the field of image processing technology.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image enhancement method, computer device, and storage medium based on information divide-and-conquer. Background Technology

[0002] When taking photographs, low-light conditions often limit the shooting environment or the photographer's unfamiliarity with photographic techniques, resulting in dark images. To improve the quality of these dark images, enhancement processing is necessary. Current dark image enhancement techniques include histogram processing and gamma correction, which work by analyzing the distribution of optical parameters in the dark image to enhance individual pixels. However, because dark images inherently lack information compared to normal images, current dark image enhancement techniques are prone to imbalances, leading to poor visual quality in the enhanced image. For example, histogram-based dark image enhancement techniques can cause contrast degradation due to grayscale shifts to higher levels.

[0003] To address the limitations of dark image enhancement techniques based on analyzing the distribution patterns of optical parameters in dark images, artificial intelligence-based dark image enhancement techniques have been proposed. For example, researchers first applied neural networks to dark image enhancement tasks with LLNet, and subsequently, more and more deep learning-based neural networks emerged, with models based on complex architectures such as CNN and Transformer being proposed. Researchers proposed the SNR-LLIE-Net network, which guides the Transformer structure for dark image enhancement by extracting the signal-to-noise ratio (SNR) map of the image. Researchers also proposed the LLFlow enhancement network, composed of a conditional encoder and a reversible model. The features extracted by the conditional encoder guide the reversible model to establish a mapping between dark images and normally lit images. These methods prioritize powerful enhancement performance while neglecting the lightweight nature of the enhancement model. Meanwhile, other methods prioritize operational efficiency and lower computational cost. Researchers designed a lightweight curve estimation network, ZeroDCE, to perform dark image enhancement by dynamically adjusting the input image pixel by pixel. Researchers proposed a weight-sharing self-calibration module (SCI) to reduce computational costs. These methods have high operational efficiency, but their performance still lags behind some methods using highly complex models.

[0004] Terminology Explanation:

[0005] Wavelet transform technology, first proposed by Mallat, S., is an effective and intuitive method for image information decomposition and analysis, capable of fully extracting image content and texture information. Liu, Pengju et al. proposed a convolutional neural network model combining multi-level wavelet transforms. This method expands the receptive field by replacing traditional pooling and upsampling operations with forward and backward wavelet transforms, thereby improving image super-resolution performance. However, in current methods incorporating wavelet transforms, the number of transforms is mostly fixed, which limits performance when processing images with high content complexity, while easily leading to redundant computation when processing images with low content complexity.

[0006] DWT: Abbreviation for Discrete Wavelet Transform, representing forward discrete wavelet transform;

[0007] IWT: an abbreviation for Inverse Discrete Wavelet Transform. Summary of the Invention

[0008] In view of the poor performance and other technical problems of current artificial intelligence-based dark image enhancement technologies, the purpose of this invention is to provide an image enhancement method, core network, computer device and storage medium based on information divide-and-conquer.

[0009] On one hand, embodiments of the present invention include an image enhancement method based on information divide-and-conquer, the image enhancement method based on information divide-and-conquer includes the following steps:

[0010] Obtain the image to be processed;

[0011] The image to be processed is decomposed into multiple image blocks;

[0012] An image block classification algorithm based on the gray-level co-occurrence matrix entropy of image blocks is used to classify each image block and obtain the category label corresponding to each image block.

[0013] Based on the early termination mechanism, each image block is enhanced according to its respective category label to obtain the enhancement feature information corresponding to each image block.

[0014] Based on feature skip connections, the enhanced feature information is aggregated to obtain the enhanced result image.

[0015] Further, the step of decomposing the image to be processed into multiple image blocks includes:

[0016] The image to be processed is uniformly decomposed to obtain multiple image blocks; each image block has the same size.

[0017] Furthermore, the image patch classification algorithm based on the gray-level co-occurrence matrix entropy of image patches is used to classify each of the image patches and obtain the category label corresponding to each of the image patches, including:

[0018] Calculate the gray-level co-occurrence matrix entropy for each of the training image blocks;

[0019] A Gaussian model is established based on the gray-level co-occurrence matrix entropy corresponding to each image block;

[0020] Based on the Gaussian model and the gray-level co-occurrence matrix entropy, calculate the probability density function value corresponding to any image block;

[0021] The category label corresponding to the image patch is determined based on the probability density function value.

[0022] Furthermore, the formula used to calculate the probability density function value corresponding to any image patch based on the Gaussian model and the gray-level co-occurrence matrix entropy includes:

[0023]

[0024]

[0025] in, The probability density function value is... The mean value of the entropy of the gray-level co-occurrence matrix is ​​given. The entropy of the gray-level co-occurrence matrix is... and represents the parameters in the Gaussian model.

[0026] Further, determining the category label corresponding to the image patch based on the probability density function value includes:

[0027] Define several category labels; each category label corresponds to its respective level of complexity;

[0028] Based on the probability density function value corresponding to the image patch, the corresponding category label is selected; the complexity of the category label is positively correlated with the probability density function value.

[0029] The image blocks are labeled with the selected category tags.

[0030] Furthermore, the method based on the early termination mechanism, which enhances each image patch according to each category label to obtain the enhancement feature information corresponding to each image patch, includes:

[0031] An enhancement stage network is established; the enhancement stage network includes a first Transformer module, a second Transformer module, a third Transformer module, a first forward discrete wavelet transform module, a second forward discrete wavelet transform module, a first inverse discrete wavelet transform module, a second inverse discrete wavelet transform module, a third inverse discrete wavelet transform module, a first residual module, a second residual module, and a third residual module;

[0032] For any of the image blocks:

[0033] When the category label corresponding to the image patch represents the first category, the test feature information corresponding to the image patch is input into the first Transformer module, and the output result of the first Transformer module is obtained as the enhanced feature information;

[0034] When the category label corresponding to the image patch represents the second category, the test feature information corresponding to the image patch is input into the first Transformer module. The output of the first Transformer module is decomposed into high-frequency and low-frequency components by the first forward discrete wavelet transform module. The high-frequency components are processed by the first residual module, and the low-frequency components are processed by the second Transformer module. The output results of the first residual module and the second Transformer module are processed by the first inverse discrete wavelet transform module to output the enhanced feature information.

[0035] When the category label corresponding to the image patch represents the third category, the test feature information corresponding to the image patch is input to the first Transformer module. The output of the first Transformer module is decomposed by the first forward discrete wavelet transform module to obtain initial high-frequency components and initial low-frequency components. The initial high-frequency components are processed by the first residual module, and the initial low-frequency components are processed by the second Transformer module. The output of the first residual module is input to the third residual module for enhancement. The output of the second Transformer module is decomposed by the second forward discrete wavelet transform module to obtain enhanced high-frequency components and enhanced low-frequency components. The enhanced high-frequency components are processed by the second residual module and then processed by the second inverse discrete wavelet transform module. The enhanced low-frequency components are processed by the third Transformer module. The output of the second inverse discrete wavelet transform module and the third Transformer module are processed by the third inverse discrete wavelet transform module to output the enhanced feature information.

[0036] Furthermore, the step of enhancing each image patch based on the early termination mechanism according to each category label to obtain the enhancement feature information corresponding to each image patch also includes:

[0037] Define a loss function; the loss function is: ;in, For hyperparameters, , The test feature information, For the enhanced feature information, , This indicates that the VGG19 network was used to extract features. , This indicates the position of each pixel in the enhanced feature information. Represents the normalized pixel points RGB mean The operator for calculating cosine similarity. , express and covariance, express standard deviation express standard deviation Represents a constant;

[0038] The enhancement stage network is trained using the loss function.

[0039] Furthermore, the aggregation of the enhanced feature information based on feature skip connections to obtain the enhanced result image includes:

[0040] The test feature information corresponding to each of the image blocks is aggregated with the enhancement feature information to obtain the enhancement result image.

[0041] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory for storing at least one program, and the processor for loading at least one program to execute an information divide-and-conquer-based image enhancement method according to the embodiments.

[0042] On the other hand, embodiments of the present invention also include a storage medium storing a processor-executable program, which, when executed by a processor, is used to perform an information divide-and-conquer-based image enhancement method in the embodiments.

[0043] The beneficial effects of the present invention are as follows: The image enhancement method based on information divide-and-conquer in the embodiments uses an information divide-and-conquer strategy. By using an image block classification algorithm based on gray-level co-occurrence matrix entropy, the image blocks are divided into multiple categories according to their information content. Under the guidance of the early termination mechanism, network structures of different complexities are used to enhance the image blocks with different information content accordingly, thereby achieving good performance while ensuring low computational cost of the model. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the steps of the information divide-and-conquer-based image enhancement method in the embodiment;

[0045] Figure 2 This is a schematic diagram of the dark graph enhancement network used in the embodiment;

[0046] Figure 3 This is a schematic diagram illustrating the principle of the enhancement phase in the embodiment;

[0047] Figure 4 This is a schematic diagram illustrating the principle of the aggregation stage in the embodiment. Detailed Implementation

[0048] In this embodiment, refer to Figure 1 The information divide-and-conquer-based image enhancement method includes the following steps:

[0049] S1. Obtain the image to be processed;

[0050] S2. Decompose the image to be processed into multiple image blocks;

[0051] S3. Use an image block classification algorithm based on the gray-level co-occurrence matrix entropy of image blocks to classify each image block and obtain the category label corresponding to each image block;

[0052] S4. Based on the early termination mechanism, each image patch is enhanced according to its category label to obtain the enhancement feature information corresponding to each image patch;

[0053] S5. Based on feature skip connections, aggregate the enhanced feature information to obtain the enhanced result image.

[0054] In this embodiment, steps S1-S5 can be divided into a preprocessing stage, an enhancement stage, and an aggregation stage. Specifically, steps S1-S3 are steps in the preprocessing stage, step S4 is a step in the enhancement stage, and step S5 is a step in the aggregation stage.

[0055] Reference Figure 2The preprocessing, enhancement, and aggregation stages are each executed using corresponding networks. Specifically, the network used for the preprocessing stage includes one convolutional module and two residual modules; the network used for the enhancement stage includes three Transformer modules, three residual modules, two DWT modules, and three IWT modules; and the network used for the aggregation stage includes two residual modules and two fusion modules.

[0056] Figure 2 The principle behind the three stages shown is as follows: First, the preprocessing stage is responsible for decomposing the input image (the image to be processed) into image patches of the same size, classifying each image patch according to a classification algorithm, and encoding the image patch into features. The enhancement stage fully combines the divide-and-conquer strategy and the early termination mechanism. Based on the image patch category labels obtained in the preprocessing stage, different network structures are used to process image patches of different categories. That is, more complex image patches are enhanced using more complex network structures. The aggregation stage uses feature skip connections to use a fusion module to promote information exchange between the encoded features and the enhanced features, and finally outputs the enhanced result.

[0057] Reference Figure 2 In the preprocessing stage, when performing step S1, the image to be enhanced is obtained, wherein the image to be enhanced may be a dark image obtained under conditions such as night scene.

[0058] Reference Figure 2 In the preprocessing stage, during step S2, the image to be processed is uniformly decomposed into multiple image blocks of the same size. In this embodiment, each image block is 64×64 pixels.

[0059] To accurately classify image patches based on complexity, this embodiment proposes an image patch classification algorithm based on the gray-level co-occurrence matrix entropy of image patches. In step S3, which involves using the image patch classification algorithm based on the gray-level co-occurrence matrix entropy of image patches to classify each image patch and obtain the corresponding category label for each image patch, the following steps can be performed:

[0060] S301. Calculate the gray-level co-occurrence matrix entropy for each of the training image blocks;

[0061] S302. Establish a Gaussian model based on the gray-level co-occurrence matrix entropy corresponding to each image block;

[0062] S303. Calculate the gray-level co-occurrence matrix entropy based on each test image block;

[0063] S304. Calculate the probability density function value corresponding to the test image patch based on the Gaussian model and the gray-level co-occurrence matrix entropy;

[0064] S305. Determine the category label corresponding to the image patch based on the probability density function value.

[0065] In step S301, obtain Information of each training image patch And calculate the entropy of the gray-level co-occurrence matrix according to the following formula:

[0066]

[0067] in, This is a function for calculating the entropy of the gray-level co-occurrence matrix.

[0068] In step S302, image patch information category labels are first defined. For example, any image patch may correspond to a first category, a second category, or a third category label. The first category label indicates that the content of the image patch information (corresponding image complexity) belongs to the "simple" category; the second category label indicates that the content of the training image patch information (corresponding image complexity) belongs to the "medium" category; and the third category label indicates that the content of the training image patch information (corresponding image complexity) belongs to the "hard" category. Then, a Gaussian model is established according to the following formula:

[0069] , , , , ,

[0070] in, It is a statistical container for the entropy of the gray-level co-occurrence matrix. ,and Initially empty, through The mean and variance of the three categories of Gaussian models were calculated based on the training feature information. , , , , , wait.

[0071] In step S303, assume there are a total of Image block information Then, set ,for ,implement:

[0072] = +

[0073] After traversal is complete Then, execute:

[0074] = .

[0075] for ,implement:

[0076]

[0077] in The category labels representing the corresponding training feature information are the first category, the second category, and the third category, respectively.

[0078] After traversal is complete Calculated Then, traverse ,implement:

[0079]

[0080] Thus, the first The gray-level co-occurrence matrix entropy corresponding to each image patch (corresponding to the test feature information) .

[0081] In step S304, based on the parameters such as mean and variance in the Gaussian model, and the entropy of the gray-level co-occurrence matrix... The calculation is performed according to the following formula:

[0082]

[0083] Thus, the first... The image patch (corresponding to the test feature information) corresponds to the first image patch in the... Probability density function values ​​under each category label .

[0084] In step S305, for the first Image patches, based on three probability density function values The size of the function value determines the category label for the first, second, or third category, based on the largest function value. Each image block is labeled.

[0085] By executing steps S301-S305, image blocks can be accurately classified based on the gray-level co-occurrence matrix entropy of the image blocks and the complexity of their content.

[0086] In this embodiment, when performing step S4 in the enhancement stage, which is based on the early termination mechanism to enhance each image patch according to each category label and obtain the enhancement feature information corresponding to each image patch, the following steps can be performed:

[0087] S401. Establish the enhanced phase network;

[0088] For any image patch, perform the following steps:

[0089] S40201.Reference Figure 3 In part (a), when the category label corresponding to the image patch represents the first category, the test feature information corresponding to the image patch is input into the first Transformer module T1, and the output result of the first Transformer module T1 is obtained as the enhanced feature information;

[0090] S40202.Reference Figure 3 In part (b), when the category label corresponding to the image patch represents the second category, the test feature information corresponding to the image patch is input into the first Transformer module T1. The output of the first Transformer module T1 is decomposed into high-frequency components and low-frequency components by the first forward discrete wavelet transform module DWT1. The high-frequency components are processed by the first residual module R1, and the low-frequency components are processed by the second Transformer module T2. The output results of the first residual module R1 and the second Transformer module T2 are processed by the first inverse discrete wavelet transform module IWT1 to output the enhanced feature information.

[0091] S40203.Reference Figure 3 In part (c), when the category label corresponding to the image patch represents the third category, the test feature information corresponding to the image patch is input into the first Transformer module T1. The output of the first Transformer module T1 is decomposed by the first forward discrete wavelet transform module DWT1 to obtain the initial high-frequency component and the initial low-frequency component. The initial high-frequency component is processed by the first residual module R1, and the initial low-frequency component is processed by the second Transformer module T2. The output of the first residual module R1 is input to the third residual module R3 for enhancement. The output of the second Transformer module T2 is decomposed by the second forward discrete wavelet transform module DWT2 to obtain the enhanced high-frequency component and the enhanced low-frequency component. The enhanced high-frequency component is processed by the second residual module R2 and then processed by the second inverse discrete wavelet transform module IWT2. The enhanced low-frequency component is processed by the third Transformer module T3. The third inverse discrete wavelet transform module IWT3 processes the outputs of the second inverse discrete wavelet transform module IWT2 and the third Transformer module T3 to output the enhanced feature information.

[0092] In step S401, refer to Figure 2 and Figure 3The enhancement stage network includes a first Transformer module T1, a second Transformer module T2, a third Transformer module T3, a first forward discrete wavelet transform module DWT1, a second forward discrete wavelet transform module DWT2, a first inverse discrete wavelet transform module IWT1, a second inverse discrete wavelet transform module IWT2, a third inverse discrete wavelet transform module IWT3, a first residual module R1, a second residual module R2, and a third residual module R3.

[0093] By executing steps S40201-S40203 and combining them with the early termination mechanism, different types of image blocks will be processed using network structures of corresponding complexity, and will undergo different numbers of wavelet transforms, thereby achieving image enhancement based on information divide and conquer.

[0094] In this embodiment, when performing step S5 of the aggregation stage, which is to aggregate the enhanced feature information based on feature skip connections to obtain the enhanced result image, the test feature information and enhanced feature information corresponding to each image block are aggregated respectively to obtain the enhanced result image.

[0095] The principle of step S5 is as follows: the aggregation stage function recovers the image from the enhanced features; when the enhanced feature information obtained from step S4 enters the aggregation stage, it is first stitched together to restore the original size, and then further processed by the residual module; in order to better reconstruct the details of the image, refer to Figure 2 and Figure 4 Step S5 uses feature skip connections to connect the features (test feature information) output from the preprocessing stage to the aggregation stage.

[0096] Step S5 uses a network that combines a fusion module with a channel attention mechanism and is inserted between residual modules. The structure of this network is as follows: Figure 4 As shown, after a series of processing steps in the aggregation stage, the final output image is enhanced, which is beneficial for further promoting feature fusion.

[0097] for Figure 2 The dark image enhancement network, consisting of the networks used in the preprocessing stage, the enhancement stage, and the aggregation stage, can be trained using a joint loss function. Perceived loss item Weighted color loss Image contrast loss term .

[0098] Regarding image content loss The definition is as follows:

[0099]

[0100] in , These represent the output enhanced image (test feature information) and the reference image (enhanced feature information), respectively.

[0101] Perceived loss term The definition is as follows:

[0102]

[0103] in This indicates that a pre-trained VGG19 network is used to extract features. Additionally, the weighted color loss term is defined as follows:

[0104]

[0105] in This indicates the position of each pixel in the enhanced image. Represents the normalized pixel points RGB mean Operators for calculating cosine similarity.

[0106] The definition of contrast loss is as follows:

[0107]

[0108] in This represents the covariance between the enhanced image and the reference image. and These represent the standard deviations of the enhanced image and the reference image, respectively. This represents a very small constant, for example, it can take the value 0.01.

[0109] The total loss function is expressed as:

[0110]

[0111] in These are hyperparameters, and in this method they are set to... =20, =0.5 and =0.5. This method is trained for 10,000 rounds on the dark map dataset.

[0112] Experimental comparisons of enhancement effects in steps S1-S6 confirm that the information divide-and-conquer-based image enhancement method in this embodiment improves the PSNR by approximately 1.6 dB on the commonly used dark image enhancement dataset LOL, and also improves the SSIM and LPIPS metrics by 0.01 and 0.009, respectively. In terms of computation, the information divide-and-conquer-based image enhancement method in this embodiment requires less than 25 GB of computation, significantly lower than state-of-the-art methods (approximately 29 GB). Furthermore, the model has only 1.4 M parameters, achieving lightweight design, fully demonstrating the effectiveness of the information divide-and-conquer-based image enhancement method in this embodiment. The information divide-and-conquer-based image enhancement method in this embodiment can be applied to image illumination enhancement tasks (including dark image enhancement and night scene image enhancement tasks), and can also be extended to high-level image vision tasks.

[0113] A computer program that executes the information divide-and-conquer-based image enhancement method in this embodiment can be written into a computer device or storage medium. When the computer program is read out and run, the information divide-and-conquer-based image enhancement method in this embodiment is executed, thereby achieving the same technical effect as the information divide-and-conquer-based image enhancement method in the embodiment.

[0114] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the components of this disclosure in the accompanying drawings. The singular forms "a," "an," and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.

[0115] It should be understood that although the terms first, second, third, etc., may be used to describe various elements in this disclosure, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. The use of any and all instances or exemplary language (“e.g.,” “such as,” etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of the invention.

[0116] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0117] Furthermore, the procedures described in this embodiment can be performed in any suitable order unless otherwise indicated by this embodiment or clearly contradicted by the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. A computer program includes multiple instructions executable by one or more processors.

[0118] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention of this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques of the invention, the invention also includes the computer itself.

[0119] A computer program can be applied to input data to perform the functions of this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.

[0120] The above are merely preferred embodiments of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. An image enhancement method based on information divide-and-conquer, characterized in that, The information divide-and-conquer based image enhancement method includes: Obtain the image to be processed; The image to be processed is decomposed into multiple image blocks; An image block classification algorithm based on the gray-level co-occurrence matrix entropy of image blocks is used to classify each image block and obtain the category label corresponding to each image block. Based on the early termination mechanism, each image block is enhanced according to its respective category label to obtain the enhancement feature information corresponding to each image block. Based on feature skip connections, the test feature information corresponding to each of the image blocks is aggregated with the enhancement feature information to obtain the enhancement result image; The method based on early termination enhances each image patch according to its respective category label to obtain enhancement feature information corresponding to each image patch, including: Establish an enhanced phase network; For any of the image blocks: When the category label corresponding to the image patch represents the first category, the test feature information output from the preprocessing stage corresponding to the image patch is input into the first Transformer module, and the output result of the first Transformer module is obtained as the enhanced feature information; When the category label corresponding to the image patch represents the second category, the test feature information corresponding to the image patch is input into the first Transformer module. The output of the first Transformer module is decomposed into high-frequency and low-frequency components by the first forward discrete wavelet transform module. The high-frequency components are processed by the first residual module, and the low-frequency components are processed by the second Transformer module. The output results of the first residual module and the second Transformer module are processed by the first inverse discrete wavelet transform module to output the enhanced feature information. When the category label corresponding to the image patch represents the third category, the test feature information corresponding to the image patch is input to the first Transformer module. The output of the first Transformer module is decomposed by the first forward discrete wavelet transform module to obtain initial high-frequency components and initial low-frequency components. The initial high-frequency components are processed by the first residual module, and the initial low-frequency components are processed by the second Transformer module. The output of the first residual module is input to the third residual module for enhancement. The output of the second Transformer module is decomposed by the second forward discrete wavelet transform module to obtain enhanced high-frequency components and enhanced low-frequency components. The enhanced high-frequency components are processed by the second residual module and then processed by the second inverse discrete wavelet transform module. The enhanced low-frequency components are processed by the third Transformer module. The output of the second inverse discrete wavelet transform module and the third Transformer module are processed by the third inverse discrete wavelet transform module to output the enhanced feature information.

2. The image enhancement method based on information divide-and-conquer according to claim 1, characterized in that, The step of decomposing the image to be processed into multiple image blocks includes: The image to be processed is uniformly decomposed to obtain multiple image blocks; each image block has the same size.

3. The image enhancement method based on information divide-and-conquer according to claim 1, characterized in that, The image patch classification algorithm based on the gray-level co-occurrence matrix entropy of image patches is used to classify each image patch and obtain the category label corresponding to each image patch, including: Calculate the gray-level co-occurrence matrix entropy for each training image patch; A Gaussian model is established based on the gray-level co-occurrence matrix entropy corresponding to each training image block; Calculate the gray-level co-occurrence matrix entropy based on each test image block; Based on the Gaussian model and the gray-level co-occurrence matrix entropy corresponding to the test image block, calculate the probability density function value corresponding to any test image block; The category label corresponding to the test image block is determined based on the probability density function value.

4. The image enhancement method based on information divide-and-conquer according to claim 3, characterized in that, The formula used to calculate the probability density function value corresponding to any test image patch based on the Gaussian model and the gray-level co-occurrence matrix entropy corresponding to the test image patch includes: in, For the first The test feature information corresponding to the test image block in the th ... The probability density function values ​​under each category label The mean value of the gray-level co-occurrence matrix entropy corresponding to the test image block. The gray-level co-occurrence matrix entropy corresponding to the test image block. and Let be the parameters in the Gaussian model, and α be the hyperparameter.

5. The image enhancement method based on information divide-and-conquer according to claim 4, characterized in that, Determining the category label corresponding to the image patch based on the probability density function value includes: Define several category labels; each category label corresponds to its respective level of complexity; Based on the probability density function value corresponding to the image patch, the corresponding category label is selected; the complexity of the category label is positively correlated with the probability density function value. The image blocks are labeled with the selected category tags.

6. The image enhancement method based on information divide-and-conquer according to claim 1, characterized in that, The method based on the early termination mechanism, which enhances each image patch according to each category label to obtain the enhancement feature information corresponding to each image patch, further includes: Define a loss function; the loss function is: ;in, For hyperparameters, , The test feature information, For the enhanced feature information, , This indicates that the VGG19 network was used to extract features. , This indicates the position of each pixel in the enhanced feature information. Represents the normalized pixel points RGB mean The operator for calculating cosine similarity. , express and covariance, express standard deviation express standard deviation Represents a constant; The enhancement stage network is trained using the loss function.

7. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store at least one program, and the processor being used to load at least one program to execute the information divide-and-conquer-based image enhancement method of any one of claims 1-6.

8. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the information divide-and-conquer-based image enhancement method of any one of claims 1-6.

Images