A low-illumination image enhancement method, system, storage medium and device

Abstract

The application discloses a low-illumination image enhancement method, system, storage medium and equipment, relates to the technical field of image processing, and the method is more detailed in dividing the noise in the low-illumination image, according to the relationship between the noise and the image signal, is divided into signal related noise and signal independent noise, through the subdivision, different types of noise can be more fully processed, so that the problem of denoising amplification is avoided. And in order to avoid the interference of noise on Retinex decomposition, different from the processing flow of other methods, the method shown in the application proposes a layer-by-layer decomposition network based on the constructed mixed related noise separation low-illumination image model and Retinex theory. The network layer by layer decomposes the input image into a noise image, a clean image, a reflection component and an illumination component, which helps the network to improve the accuracy of denoising and Retinex decomposition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a low-illumination image enhancement method and system, a storage medium and an equipment. BACKGROUND

[0002] Low-illumination images usually have problems such as low illumination, much noise and loss of detail information, which will affect the visual effect of the image and subsequent applications. Specifically, first, low illumination and much noise will interfere with the visual perception of the human eye. The image after enhancing the illumination and removing noise will be smoother and more natural, which helps to improve the visual experience of users, especially in scenarios where the details of the image need to be carefully observed. Second, in image analysis, processing or recognition applications, the existence of low illumination and noise will interfere with the performance of the algorithm. Through image enhancement, the accuracy and efficiency of these algorithms can be significantly improved.

[0003] Among them, the supervised low-illumination image enhancement method needs a large number of low-illumination / normal-illumination image pair data to train the model. It is difficult and costly to obtain such data. Moreover, the quality of the data will affect the training effect and accuracy of the model. In addition, the generalization of the model may also be limited. When processing data that is significantly different from the training data, the results will usually have the following problems:

[0004] 1. Flexibility and local under / over exposure: A histogram graphically represents the distribution of pixel intensities in an image. Histogram equalization-based methods achieve enhancement by stretching the dynamic range of the image. However, histogram equalization-based methods lack flexibility and are difficult to adjust locally, which can cause over / under exposure and noise amplification in the enhanced image. Traditional Retinex-based methods rely on manually designed priors, so they cannot adapt to multiple different scenarios at the same time.

[0005] 2. Noise amplification and color shift: With the development of deep learning, many deep learning-based low-illumination enhancement methods have emerged in recent years. Many researchers combine Retinex theory with deep learning methods. These methods usually first perform Retinex decomposition on the input image, process the reflection component and the illumination component respectively, and then obtain the enhanced image by reconstruction. Such methods have good enhancement effect, but do not explicitly distinguish and process the defects existing in the reflection component, resulting in noise and color shift in the enhanced image.

[0006] 3. Data dependency and data set construction: supervised low-light image enhancement methods rely on low / normal light image pairs, and there are certain difficulties in data collection. Generally, low-light and normal-light image pairs need to be collected under different lighting conditions (night, rainy day, indoor environment, etc.). The collection process not only consumes time and effort, but also may be affected by various factors such as weather, equipment, shooting angle, etc., resulting in uneven data quality. SUMMARY

[0007] In view of the deficiencies of the prior art, the purpose of the present application is to provide a low-light image enhancement method, system, storage medium and device, which aims to solve the above problems described in the background art.

[0008] The first aspect of the present application is to provide a low-light image enhancement method, the method comprising:

[0009] obtaining a denoised low-light image, converting the color space corresponding to the low-light image from RGB color space to HSV color space, and extracting the brightness information of the low-light image;

[0010] According to the brightness information, the low-light image is divided into dark and bright regions according to a preset brightness threshold;

[0011] The area proportion of the dark region and the bright region in the low-light image is calculated respectively, the overall brightness of the low-light image is estimated, and the illumination component corresponding to the dark region is outputted;

[0012] The illumination component corresponding to the dark region is adjusted by using a gamma correction method, and the Gamma value is calculated according to the overall brightness, so as to realize the adaptive correction and image enhancement of the low-light image, and obtain a target image.

[0013] According to one aspect of the above technical solution, the step of calculating the area proportion of the dark region and the bright region in the low-light image, estimating the overall brightness of the low-light image, and outputting the illumination component corresponding to the dark region comprises:

[0014] obtaining the image area of the low-light image, and identifying the dark area of the dark region and the bright area of the bright region in the low-light image;

[0015] According to the image area of the low-light image and the dark area of the dark region and the bright area of the bright region, the area proportion of the dark region and the bright region in the low-light image is calculated respectively;

[0016] According to the area proportion of the dark region and the bright region in the low-illumination image, the overall brightness of the low-illumination image is estimated, and the illumination component corresponding to the dark region is output.

[0017] According to an aspect of the above technical solution, the expression for estimating the overall brightness of the low-illumination image is:

[0018] ;

[0019] In the formula, B is the overall brightness, is the dark area of the dark region, is the bright area of the bright region, θ is the brightness threshold, is the average brightness of the dark region, is the average brightness of the bright region, L is the illumination component corresponding to the dark region, H is the illumination component corresponding to the bright region.

[0020] According to an aspect of the above technical solution, the illumination component corresponding to the dark region is adjusted using a gamma correction method, and the Gamma value is calculated according to the overall brightness, so as to realize the steps of adaptive correction and image enhancement of the low-illumination image. In the step of adjusting the illumination component using a gamma correction method, the expression is:

[0021] ;

[0022] In the formula, is the compensation coefficient, is the gamma coefficient, and ;

[0023] Wherein, when the smaller the value, the higher the image brightness, and if the larger the value, the lower the image brightness.

[0024] According to an aspect of the above technical solution, the method further comprises:

[0025] Based on the preset Retinex model, the mutually independent illumination components obtained by decomposition are inversely reconstructed, the target illumination component after gamma correction is multiplied by the preset reflection component pixel by pixel, and an enhanced target image is obtained.

[0026] According to an aspect of the above technical solution, the expression for multiplying the target illumination component after gamma correction by the preset reflection component pixel by pixel to obtain the enhanced target image is:

[0027] ;

[0028] wherein, is a target image, R is a reflection component, is a target illumination component.

[0029] A second aspect of the present application provides a low-illumination image enhancement system, applied to the method in the above technical solution, and the system comprises:

[0030] a color space conversion module, configured to acquire a low-illumination image after noise reduction, convert a color space corresponding to the low-illumination image from an RGB color space to an HSV color space, and extract luminance information of the low-illumination image;

[0031] a luminance region division module, configured to divide the low-illumination image into a dark region and a bright region according to a preset luminance threshold value based on the luminance information;

[0032] an image parameter calculation module, configured to calculate area proportions of the dark region and the bright region in the low-illumination image respectively, estimate an overall brightness of the low-illumination image, and output an illumination component corresponding to the dark region;

[0033] an image correction and enhancement module, configured to adjust the illumination component corresponding to the dark region in a gamma correction manner, and calculate a Gamma value based on the overall brightness, so as to realize adaptive correction and image enhancement of the low-illumination image.

[0034] According to an aspect of the above technical solution, the luminance region division module is specifically configured to:

[0035] acquire an image area of the low-illumination image, and identify a dark region area of the dark region and a bright region area of the bright region in the low-illumination image;

[0036] calculate the area proportions of the dark region and the bright region in the low-illumination image respectively based on the image area of the low-illumination image and the dark region area of the dark region and the bright region area of the bright region;

[0037] estimate the overall brightness of the low-illumination image based on the area proportions of the dark region and the bright region in the low-illumination image, and output the illumination component corresponding to the dark region.

[0038] A third aspect of the present application provides a readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the method in the above technical solution.

[0039] The fourth aspect of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the above technical solutions when executing the computer program.

[0040] Compared with the prior art, the low-illumination image enhancement method, system, storage medium and device provided by the present application have the beneficial effects that:

[0041] The method provided by the present application can more fully process different types of noise by more detailed division of noise in a low-illumination image, and the noise is divided into signal-related noise and signal-unrelated noise according to the relationship between the noise and the image signal, thereby avoiding the problem of noise amplification. In order to avoid the interference of noise on Retinex decomposition, the method provided in the present application proposes a layer-by-layer decomposition network based on the constructed mixed related noise separation low-illumination image model and the Retinex theory, which helps to improve the accuracy of noise removal and Retinex decomposition. BRIEF DESCRIPTION OF DRAWINGS

[0042] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 A flowchart of a low-illumination image enhancement method in an embodiment of the present application;

[0044] Figure 2 A structure block diagram of a low-illumination image enhancement system in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the objectives, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, the present application can be realized in many different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0046] It should be noted that when an element is referred to as being "fixedly attached" to another element, it can be directly on the other element or there can be an intervening element. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements can be present. The terms "vertical", "horizontal", "left", "right", and the like as used herein are for purposes of illustration and description only.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0048] Embodiment one

[0049] Please refer to Figure 1 The first embodiment of the present application provides a low-illumination image enhancement method, which comprises steps S10-S40:

[0050] Step S10, obtaining a low-illumination image after noise reduction, converting the color space corresponding to the low-illumination image from RGB color space to HSV color space, and extracting the brightness information of the low-illumination image.

[0051] First of all, it should be pointed out that RGB color space and HSV color space are two commonly used color representation methods. RGB color space is based on three basic colors of red, green and blue, and produces rich and extensive colors by different degrees of superposition. In RGB color space, the value range of red, green and blue is 0-255. Among them, 0 represents the minimum intensity of the color component, and 255 represents the maximum intensity. While HSV color space uses hue (H), saturation (S) and value (V) to describe color. Hue H parameter represents color information, i.e. the position of the spectral color, which is represented by angle and the value range is 0°-360°. The value range of saturation S and brightness V is 0.0-1.0.

[0052] In this embodiment, the low-illumination image after noise reduction is converted from RGB color space to HSV color space, and the V channel is extracted to obtain the brightness information of the image.

[0053] Step S20, according to the brightness information, the low-illumination image is divided into dark region and bright region according to the preset brightness threshold.

[0054] In this embodiment, after the V channel is extracted, i.e. the brightness information of the image is obtained, a brightness threshold is first set to divide the dark region and the bright region, the region with brightness value lower than the threshold is divided into the dark region, and the region with brightness value higher than the threshold is divided into the bright region.

[0055] Step S30, respectively calculating the area proportion of the dark region and the bright region in the low-illumination image, estimating the overall brightness of the low-illumination image, and outputting the illumination component corresponding to the dark region.

[0056] In the embodiment, the step of respectively calculating the area proportion of the dark region and the bright region in the low-illumination image, estimating the overall brightness of the low-illumination image, and outputting the illumination component corresponding to the dark region comprises:

[0057] obtaining the image area of the low-illumination image, and identifying the dark region area of the dark region and the bright region area of the bright region in the low-illumination image;

[0058] According to the image area of the low-illumination image and the dark region area of the dark region and the bright region area of the bright region, respectively calculating the area proportion of the dark region and the bright region in the low-illumination image;

[0059] According to the area proportion of the dark region and the bright region in the low-illumination image, estimating the overall brightness of the low-illumination image, and outputting the illumination component corresponding to the dark region.

[0060] Wherein, the proportion of the dark region and the bright region in the whole image is calculated respectively, the area size of the dark region is , and the area size of the bright region is . According to the division of the bright region and the dark region, the overall brightness of the image is estimated , and the expression is:

[0061] ;

[0062] Wherein, is the overall brightness, is the average brightness of the dark region, is the average brightness of the bright region, is the brightness threshold, is the average brightness of the dark region, is the average brightness of the bright region, is the illumination component corresponding to the dark region, is the illumination component corresponding to the bright region.

[0063] Step S40, adjusting the illumination component corresponding to the dark region by using the gamma correction method, and calculating the Gamma value according to the overall brightness, so as to realize the adaptive correction and image enhancement of the low-illumination image, and obtaining the target image.

[0064] In this embodiment, after denoising and Retinex decomposition using a layer-by-layer decomposition network, the illuminance components corresponding to the dark areas are... Further adjustments were made to improve the image brightness.

[0065] Among them, gamma correction was used to control the degree component. Adjustments are made, and the gamma value is calculated based on the overall image brightness obtained from region-weighted estimation, thereby achieving adaptive image brightness correction to meet the enhancement needs of images with different brightness levels. The expression is as follows:

[0066] ;

[0067] In the formula, This is the compensation coefficient, which is typically 0. The gamma coefficient is, and .

[0068] Specifically, during gamma correction, the gamma value The size of the [something] directly affects the brightness of the image. When The smaller the value, the greater the increase in image brightness; conversely, if... A larger value results in a smaller increase in image brightness. This can be achieved by adjusting... The illuminance component can be adjusted to normal illuminance, bringing the overall brightness of the image to a more suitable level. In this way, the processed image not only removes noise and retains rich details and texture information, but also significantly improves brightness, providing users with a clearer and more natural visual experience.

[0069] Finally, based on the Retinex model, the independent components obtained from the decomposition are reconstructed in reverse, and the target illuminance components after gamma correction are obtained. With reflection component By performing pixel-by-pixel multiplication, the enhanced target image can be obtained. The expression is:

[0070] ;

[0071] In the formula, For the target image, For the reflection component, This represents element-wise multiplication. The target illuminance component.

[0072] The low-light image enhancement method shown in this embodiment can be applied to sub-fields such as public safety, consumer electronics, medical imaging, and autonomous driving. Specifically:

[0073] Public safety: In the night or light environment, the image captured by the surveillance camera is often difficult to identify. By enhancing the image captured by the camera, the clarity and brightness of the image can be improved, which helps to improve the safety monitoring level of public areas.

[0074] Consumer electronics: With the popularity of mobile devices, consumers are increasingly demanding high-quality images. By low-light image enhancement, the shooting effect of mobile phones in dark light environment can be improved, so that users can take clear and bright photos even in insufficient light, which helps to improve the user's photography experience and satisfaction.

[0075] Medical imaging: In endoscopy, due to limited lighting conditions, the captured images often have poor visibility, making it difficult to determine the condition. By low-light image enhancement, the brightness, contrast and details of these images can be improved, so that doctors can more accurately determine the condition and develop treatment plans.

[0076] Autonomous driving: In the night, light environment and bad weather, autonomous driving systems need to rely on cameras and sensors to perceive the surrounding environment. By enhancing the camera's picture, the safety and accuracy of the autonomous driving system can be improved.

[0077] Compared with the prior art, the low-light image enhancement method shown in the embodiment has the following beneficial effects:

[0078] The method shown in the embodiment classifies the noise in the low-light image more carefully, classifies it into signal-dependent noise and signal-independent noise according to the relationship between the noise and the image signal, and processes different types of noise more fully through this subdivision, thereby avoiding the problem of denoising amplification. And to avoid the interference of noise on Retinex decomposition, unlike other processing procedures, the method shown in the embodiment proposes a layer-by-layer decomposition network based on the constructed mixed related noise separation low-light image model and Retinex theory. The network decomposes the input image layer by layer into a noise image, a clean image, a reflection component and an illumination component, which helps the network to improve the accuracy of denoising and Retinex decomposition.

[0079] Embodiment two

[0080] The second embodiment of the application also provides a low-light image enhancement method, and the method shown in the embodiment:

[0081] Affected by lighting conditions and limited by shooting equipment, low-light images contain a large amount of noise, especially in dark areas. Existing methods usually consider low-light images as being composed of a clean image and noise However, low-light images Noise in the low-light image can be classified as signal-dependent noise and signal-independent noise according to its relationship with the signal. In order to fully remove the noise in the low-light image and achieve better enhancement effect, the noise is subdivided according to the above classification, and a corresponding mixed related noise separation low-light image model is constructed, and the expression is:

[0082]

[0083] In the formula, is signal-dependent noise, is signal-independent noise.

[0084] Low-light images often have problems such as insufficient brightness, low contrast, and large noise. Traditional low-light image enhancement methods often cannot simultaneously meet the needs of denoising, enhancing brightness and contrast, and preserving details. The neural network shown in the embodiment learns and extracts deep and nonlinear features in data in a data-driven manner, which has better flexibility and generalization. Therefore, a layer-by-layer decomposition low-light image enhancement network model needs to be constructed in the embodiment to gradually decompose the input image into a signal-independent noise image, a signal-dependent noise image, a reflection component, and an illumination component. The low-light image enhancement network system shown in the embodiment includes:

[0085] The first layer: a mixed related noise separation low-light image model. The first layer network separates the signal-independent noise from the input image, i.e., the low-light image , to remove this type of noise. By designing a self-supervised image reconstruction loss and a signal-independent noise consistency loss, the first layer network of the layer-by-layer decomposition network is jointly constrained.

[0086] Based on the constructed low-light image model, the input image, i.e., the low-light image is first decomposed by the first layer network . This layer network is composed of 2 parallel convolution branches, each branch containing 5 layers of convolution. Through the 2 convolution branches, including and , the image containing signal-dependent noise and the signal-independent noise are obtained respectively. Thus, the signal-independent noise is removed.

[0087] The second layer: based on the above image model, a self-supervised image consistency loss constraint is designed, and the second layer network of the layer-by-layer decomposition network is jointly constrained by the above image reconstruction loss. This layer network further separates the signal-dependent noise in the image, thereby obtaining a completely denoised low-light image. ​​​

[0088] where the second layer network of the layer-by-layer decomposition network is decomposed , further decomposes the image containing signal-dependent noise. With the same network structure, also consists of two parallel convolution branches. Through its and The two convolution branches respectively obtain the clean image and signal-dependent noise , so as to remove the signal-dependent noise , obtain the clean image .

[0089] The third layer: based on the standard Retinex model, design Retinex loss and reflection component consistency loss, joint constraint the Retinex decomposition network of the last layer, so as to complete the construction of the layer-by-layer decomposition low-light image enhancement network model. The network further decomposes the clean low-light image into a reflection component representing the object properties in the image and an illumination component representing the environmental illumination information.

[0090] where the Retinex model is realized based on the Retinex theory. Retinex is a composite word composed of retina and cortex. According to the Retinex theory, an image depends on incident light and the reflection of the object surface to the incident light. Therefore, the low-light image is composed of a reflection component (reflectance) and an illumination component (Illumination), and the expression is:

[0091] ;

[0092] In the formula, the reflection component represents the physical properties of the object in the image, the illumination component represents the light intensity of the shooting environment, and represents element-wise multiplication.

[0093] Specifically, the Retinex theory has good physical interpretability, so it is widely used in various low-light image enhancement methods. After denoising in the first and second layers, the clean image is obtained. Then, based on the standard Retinex model, the clean image is decomposed by . Like the two sub-networks in the first and second layers, it consists of two parallel convolution branches, through its and 2 convolution branches respectively obtain the reflection component and the illumination component .

[0094] In addition, the supervised method relies on low / normal illumination image pairs, and there are certain difficulties in data collection. The unsupervised low-light enhancement method does not rely on low / normal illumination paired image pairs, but the information available from a single image is limited. Therefore, the low-light image pair dataset is used as the training dataset in the embodiment. The dataset contains 324 scenes, each scene has 2 to 6 low-light images with different illuminations, a total of 1000 images. When training the above layer-by-layer decomposition low-light image enhancement network model, the input of the network is a low-light image pair, that is, two low-light images with different brightness in the same scene and . For scenes containing more than 2 images, the images with different illuminations are paired two by two as the input of the model.

[0095] In the embodiment, the layer-by-layer decomposition network is constrained by the self-supervised image reconstruction loss, the signal-independent noise consistency loss, the image consistency loss, the reflection component consistency loss and the Retinex loss. Specifically, the self-supervised image reconstruction loss and the signal-independent noise consistency loss are designed to jointly constrain the first layer network of the layer-by-layer decomposition network; the self-supervised image consistency loss is designed to jointly constrain the second layer network of the layer-by-layer decomposition network with the image reconstruction loss; the Retinex loss and the reflection component consistency loss are designed to jointly constrain the Retinex decomposition network at the end. Among them:

[0096] a. Image reconstruction loss, expression is:

[0097] ;

[0098] According to the constructed image model, the three components , and obtained by decomposition should be able to reconstruct the input image , that is, the low-light image . In addition, according to the image model, the signal-dependent noise can also be calculated from the input image , the clean image and the signal-independent noise .

[0099] Among them, the second term is to make the decomposition of the second layer to more consistent with the proposed mixed related noise separation low-light image model, that is, mainly for the decomposition of . At this time, it can be considered that The first term of the image reconstruction loss and the following image consistency loss have been accurately decomposed. The main role of the second term of the image reconstruction loss is to further accurately decompose . Therefore, The part of the network parameters related should not be updated with the calculation of the second term of the reconstruction loss, so stopgrad is used to make the decomposition part of the network ( ) not participate in gradient calculation, avoid affecting the final decomposition effect, make the decomposition result more accurate.

[0100] b. Signal-independent noise consistency loss, expression:

[0101] ;

[0102] Where the signal-independent noise is independent of the image signal, mainly produced when the camera scans and processes the image during shooting. Therefore, the signal-independent noise of the image pair of the same scene taken by the same camera should be consistent.

[0103] c. Image consistency loss, expression:

[0104] ;

[0105] Where the image content should not change before and after denoising. To ensure the consistency of the image content before and after denoising, the clean image is constrained.

[0106] d. Reflection component consistency loss, expression:

[0107] ;

[0108] Where the input image is a low-illumination image pair of the same scene content and different illuminance. Therefore, the reflection component , representing the physical properties of the same object, its signal strength and gradient should remain consistent.

[0109] e. Retinex loss, expression:

[0110] ;

[0111] Where, according to the Retinex theory, the decomposed and can be reconstructed to obtain the input clean image . And, from the clean image and ​ To ensure the accuracy of Retinex decomposition, stopgrad is also used here. The illumination map is the maximum value of the three channels of the input image, and the illumination component should be approximately equal to it. In addition, the illumination component represents the illumination information of the environment and does not contain the texture characteristics of the object, so it should be as smooth as possible.

[0112] Embodiment Three

[0113] Please refer to Figure 2 , the third embodiment of the present application provides a low-illumination image enhancement system, which is applied to the method described in the above embodiments, and the system comprises:

[0114] A color space conversion module 10 is configured to obtain a low-illumination image after noise reduction, convert the color space corresponding to the low-illumination image from an RGB color space to an HSV color space, and extract the brightness information of the low-illumination image.

[0115] A brightness area division module 20 is configured to divide the low-illumination image into a dark area and a bright area according to a preset brightness threshold value based on the brightness information.

[0116] An image parameter calculation module 30 is configured to calculate the area proportion of the dark area and the bright area in the low-illumination image respectively, estimate the overall brightness of the low-illumination image, and output the illumination component corresponding to the dark area.

[0117] An image correction and enhancement module 40 is configured to adjust the illumination component corresponding to the dark area by using a gamma correction method, and calculate the Gamma value based on the overall brightness to realize adaptive correction and image enhancement of the low-illumination image.

[0118] In this embodiment, the brightness area division module is specifically configured to:

[0119] obtain the image area of the low-illumination image, and identify the dark area of the dark area and the bright area of the bright area in the low-illumination image.

[0120] According to the image area of the low-illumination image and the dark area of the dark area and the bright area of the bright area, the area proportion of the dark area and the bright area in the low-illumination image is calculated respectively.

[0121] According to the area proportion of the dark area and the bright area in the low-illumination image, the overall brightness of the low-illumination image is estimated, and the illumination component corresponding to the dark area is output.

[0122] Compared with the prior art, the low-illumination image enhancement system shown in the embodiment has the beneficial effects that:

[0123] The system shown in the embodiment classifies the noise in the low-illumination image more carefully, classifies it into signal-related noise and signal-unrelated noise according to the relationship between the noise and the image signal, and can more fully process different types of noise to avoid the problem of denoising amplification. In order to avoid the interference of noise on Retinex decomposition, the system shown in the embodiment proposes a layer-by-layer decomposition network based on the constructed mixed related noise separation low-illumination image model and Retinex theory, which helps the network improve the accuracy of denoising and Retinex decomposition.

[0124] Embodiment Four

[0125] The fourth embodiment of the present application provides a readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the method described in the above embodiments.

[0126] Embodiment Five

[0127] The fifth embodiment of the present application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to realize the method described in the above embodiments.

[0128] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0129] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A low-light image enhancement method, characterized by, The method comprises: obtaining a low-illumination image denoised by a layer-by-layer decomposed low-illumination image enhancement network model, converting a color space corresponding to the low-illumination image from an RGB color space to an HSV color space, and extracting brightness information of the low-illumination image; according to the brightness information, dividing the low-illumination image into a dark region and a bright region according to a preset brightness threshold; calculating area proportions of the dark region and the bright region in the low-illumination image respectively, estimating overall brightness of the low-illumination image according to region weighting, and outputting an illumination component obtained by Retinex decomposition of the clean low-illumination image after denoising; adjusting the illumination component of the clean low-illumination image by using a gamma correction method, and calculating a Gamma value according to the overall brightness of the clean low-illumination image to realize adaptive correction and image enhancement of the illumination component, so as to obtain a target illumination component; based on a preset Retinex model, reversely reconstructing the mutually independent illumination component obtained by decomposition, multiplying the target illumination component after gamma correction and the reflection component obtained by decomposition pixel by pixel to obtain an enhanced target image; wherein the layer-by-layer decomposed low-illumination image enhancement network model is constructed based on a low-illumination image model of mixed relevant noise separation, and the expression is: ； wherein is the input low-light image, is the clean low-light image, is the signal-dependent noise, is the signal-independent noise; The layer-by-layer decomposed low-illumination image enhancement network model comprises: a first layer network for decomposing a low-illumination image to obtain a low-illumination image containing signal-dependent noise and signal-independent noise, and removing the signal-independent noise; a second layer network for decomposing the low-illumination image containing signal-dependent noise again to obtain signal-dependent noise and a clean low-illumination image, and removing the signal-dependent noise; a third layer network for Retinex decomposition of the clean low-illumination image to obtain a reflection component representing object properties in the low-illumination image and an illumination component representing environmental illumination information; The first layer network, the second layer network and the third layer network each comprise two parallel convolution branches, and each convolution branch comprises five convolution layers. Wherein, the layer-by-layer decomposed low-illumination image enhancement network model is trained by using a pair of low-illumination images, the pair of low-illumination images are two low-illumination images of different illuminations in the same scene, and in the training process, the image reconstruction loss, the signal-independent noise consistency loss, the image consistency loss, the reflection component consistency loss and the Retinex loss are used for constraint; In the expression, the overall brightness of the low-illumination image is estimated according to the region weighting The expression is: ； In the formula, a dark area of a dark portion region, a bright area of a bright portion region, an average brightness of a dark portion region, an average brightness of a bright portion region, a brightness threshold value; Wherein, the expression for adjusting the illumination component of the clean low-illumination image by using the gamma correction method is: ； In the formula, is a compensation coefficient, is a gamma coefficient, and When The smaller the value, the higher the image brightness is improved, and if The larger the value, the lower the image brightness is improved; Wherein, the expression for multiplying the target illumination component after gamma correction and the reflection component obtained by decomposition pixel by pixel to obtain the enhanced target image is: ； wherein is a target image, is a reflection component, is a target illumination component.

2. The low-light image enhancement method of claim 1, wherein, The steps of calculating the area proportions of the dark region and the bright region in the low-illumination image respectively, estimating the overall brightness of the low-illumination image according to region weighting, and outputting the illumination component obtained by Retinex decomposition of the clean low-illumination image after denoising, comprise: obtain an image area of the low-illumination image, and identify a dark area and a bright area of the low-illumination image; calculate proportions of the dark area and the bright area in the low-illumination image according to the image area of the low-illumination image and the dark area and the bright area; estimate the overall brightness of the low-illumination image according to the proportions of the dark area and the bright area in the low-illumination image according to area weighting, and output an illumination component obtained by Retinex decomposition of the clean low-illumination image after denoising.

3. A low-light level image enhancement system characterized by, The system is applied to the method of any one of claims 1-2, and the system comprises: a color space conversion module configured to obtain the low-illumination image after denoising by the layer-by-layer decomposition low-illumination image enhancement network model, convert a color space corresponding to the low-illumination image from an RGB color space to an HSV color space, and extract brightness information of the low-illumination image; a brightness area division module configured to divide the low-illumination image into a dark area and a bright area according to a preset brightness threshold based on the brightness information; an image parameter calculation module configured to calculate proportions of the dark area and the bright area in the low-illumination image, estimate the overall brightness of the low-illumination image according to the proportions of the dark area and the bright area in the low-illumination image according to area weighting, and output an illumination component obtained by Retinex decomposition of the clean low-illumination image after denoising; an image correction enhancement module configured to adjust the illumination component of the clean low-illumination image in a gamma correction manner, and calculate a Gamma value based on the overall brightness of the clean low-illumination image to achieve adaptive correction and image enhancement of the illumination component, so as to obtain a target illumination component; the image correction enhancement module is further configured to perform inverse reconstruction on the independent illumination component obtained by decomposition based on a preset Retinex model, and multiply the target illumination component after gamma correction and a reflection component obtained by decomposition pixel by pixel to obtain an enhanced target image.

4. The low-light image intensification system of claim 3, wherein, The brightness area division module is specifically configured to: obtain an image area of the low-illumination image, and identify a dark area and a bright area of the low-illumination image; calculate proportions of the dark area and the bright area in the low-illumination image according to the image area of the low-illumination image and the dark area and the bright area; estimate the overall brightness of the low-illumination image according to the proportions of the dark area and the bright area in the low-illumination image according to area weighting, and output an illumination component obtained by Retinex decomposition of the clean low-illumination image after denoising.

5. A readable storage medium, having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-2.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1-2.

Images