A non-uniform low-light image enhancement method and system under zero reference samples

Abstract

The application discloses a non-uniform low-light image enhancement method and system under zero reference samples, and establishes an image decomposition network based on a robust Retinex model and deep learning; according to the image decomposition network, an input image is decomposed into an illumination map, a reflection map and a noise map; an adaptive brightness mapping function is constructed; an optimal brightness mapping curve corresponding to the illumination map is obtained; the optimal brightness mapping curve corresponding to the illumination map is used to obtain an enhanced illumination map and to obtain an enhanced image by multiplying the enhanced illumination map and the reflection map pixel by pixel; and the technical problem of poor non-uniform low-light image enhancement effect is solved. Since paired high-quality images are not required and the image brightness can be adaptively adjusted according to different illumination environments, the application can obtain natural, contrast and brightness enhanced images according to non-uniform low-light images, and has strong generalization and can adapt to different scenes.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the field of image processing technology, in particular to a non-uniform low-light image enhancement method and system under zero reference samples. BACKGROUND

[0002] Images are important carriers of information transmission, and the quality of images directly affects the visual experience and the accuracy of information transmission. However, in actual production and life, due to factors such as lighting environment and acquisition equipment, the acquired images are prone to have low and unevenly distributed light, many noise points and other characteristics, which affect the visual experience and the accuracy of visual tasks based on images, and are difficult to meet the actual application requirements. Research on non-uniform and low-light image enhancement has important application value in the fields of night monitoring, industrial detection, automatic driving and the like.

[0003] The existing non-uniform low-light image enhancement methods mainly include traditional methods and deep learning methods. The traditional methods include gray mapping method, model method and fusion method, while the deep learning methods are divided into supervised learning and unsupervised learning according to whether corresponding images under good environment are needed. The gray mapping method includes histogram equalization and gamma correction, which stretches the pixels of different brightness in the image nonlinearly to improve the overall brightness of the image. This kind of method is simple to operate, but it is prone to problems such as noise amplification and image distortion because the physical model of imaging is not considered. The model method includes atmospheric scattering model and Retinex model, which respectively utilizes atmospheric light transmission model and the color and brightness perception characteristics of human eyes to the image, and can provide better enhancement effect, but needs to manually set model parameters, has high algorithm complexity and time-consuming optimization process. The fusion method is an enhancement method of fusing multiple images with different exposure, which can have good enhancement effect because it has multiple images with different exposure, but needs image preprocessing, and the enhancement effect depends on the quality of the different images used for fusion to a certain extent. The deep learning method learns the mapping relationship between low-light images and normal light images through a neural network, or learns from low-light images to obtain enhanced images. This kind of method utilizes the powerful learning and adaptation ability of neural network, and does not need manual adjustment of parameters, but has poor interpretability and is prone to uncontrollable noise and distortion problems.

[0004] The invention patent with publication number CN111798400B discloses a non-reference low-light image enhancement method and system based on a generative adversarial network. The data set of the method is unpaired low-light and normal-light image blocks, and the generator network for low-light enhancement and the discriminator network for adversarial training are alternately trained until Nash equilibrium is reached. This method does not need paired images for supervised training, but the composition of the training data set still depends on normal-light images.

[0005] The patent with the publication number CN110232661B discloses a low-light color image enhancement method based on Retinex and a convolutional neural network. The RGB color image is decomposed into a reflection map and an illumination map through a decomposition network. The reflection map is denoised and color restored based on a reflection map recovery network. The illumination map is enhanced based on an illumination adjustment network and a user-provided illumination adjustment parameter. Finally, the enhanced result is obtained by point multiplication of the recovered reflection map. This method can achieve good enhancement effect, but it needs to manually input the illumination adjustment parameter, and the adaptability to different scenes is not high. SUMMARY

[0006] The present application provides a non-uniform low-light image enhancement method and system under zero reference samples, which solves the technical problem of poor non-uniform low-light image enhancement effect.

[0007] To solve the above technical problems, the present application provides a non-uniform low-light image enhancement method under zero reference samples, which includes:

[0008] An image decomposition network is established based on a robust Retinex model and deep learning.

[0009] According to the image decomposition network, the input image is decomposed into an illumination map, a reflection map, and a noise map.

[0010] An adaptive brightness mapping function is constructed to obtain the best brightness mapping curve corresponding to the illumination map.

[0011] The best brightness mapping curve corresponding to the illumination map is used to obtain the enhanced illumination map.

[0012] The enhanced illumination map and the reflection map are multiplied pixel by pixel to obtain an enhanced image.

[0013] Further, the robust Retinex model is specifically:

[0014]

[0015] Wherein, S represents the input image, R, I, and N represent the reflection map, the illumination map, and the noise map under the robust Retinex model, respectively, represents pixel-by-pixel multiplication.

[0016] Further, the specific formula of the loss function of the robust Retinex model is:

[0017] f dec = argmin L dec = argmin (L rec + λ1L ref + λ2L ill + λ3L n),

[0018]

[0019] wherein f dec represents a loss function of the robust Retinex model, L dec , rec , ref , ill , n respectively represent a mixed no-reference decomposition loss, a reconstruction loss, a reflection estimation loss, an illumination estimation loss and a noise estimation loss, λ1, λ2, λ3 are respectively preset reflection estimation loss weight, illumination estimation loss weight and noise estimation loss weight, S represents an input image, R, I, N respectively represent a reflection map, an illumination map and a noise map under the robust Retinex model, ||·||1 represents an L1 norm, ||·||F represents an F norm, c ∈ {r, g, b} represents RGB three channels of an image, H(·) represents histogram equalization, F represents a gradient sum in horizontal and vertical directions, and β represents a self-defined constant.

[0020] Further, the specific formula of the adaptive brightness mapping function is:

[0021]

[0022]

[0023] wherein g(·) represents a brightness mapping curve, I represents an illumination map, tanh(I) represents a hyperbolic tangent function, the value range is [-1, 1], ω controls the weight of low light enhancement, k1 and k2 are respectively used to adjust the amplitude of low light enhancement and high light suppression.

[0024] Further, the adaptive brightness mapping function is constructed to obtain the optimal brightness mapping curve corresponding to the illumination map, which comprises:

[0025] According to the adaptive brightness mapping function, the illumination enhancement network is established, the illumination enhancement network comprises seven convolutional layers and one fully connected layer, the convolutional layers are connected through ReLU activation function and maximum pooling layer, and the output of the fully connected layer is the optimal parameters ω, k1, k2.

[0026] The illumination enhancement network is trained, and the parameters of the adaptive brightness mapping function are adjusted according to the trained illumination enhancement network, so as to obtain the optimal brightness mapping curve corresponding to the illumination map.

[0027] Further, the specific formula of the objective function for training the illumination enhancement network is:

[0028] f enh ​= argmin (L E + η1L N + η2L S ),

[0029]

[0030]

[0031]

[0032] wherein f dec represents the target function of the brightness enhancement network, L E , L N , L S respectively represent the exposure control loss, the naturalness loss and the illumination smoothness loss, η1, η2 respectively represent the naturalness loss function weight and the illumination smoothness loss function weight, sign(·) represents the sign function, which takes 1 when greater than 0, -1 when less than 0, and 0 when 0, E represents the exposure, O i and I i respectively represent the brightness value of the i-th image block in the enhanced and original illumination map, Ω(i) represents the four-neighborhood space centered on the pixel i, O j , I j respectively represent the brightness value of the four-neighborhood space in the enhanced and original illumination map, represents the gradient sum in the horizontal and vertical directions of the enhanced illumination map, and β represents a self-defined constant.

[0033] Further, the specific formula for obtaining the enhanced image by pixel-wise multiplication of the enhanced illumination map and the reflection map is:

[0034]

[0035] wherein represents the enhanced image, represents the enhanced illumination map, and R represents the reflection map under the robust Retinex model.

[0036] The present application provides a non-uniform low-illumination image enhancement system under zero reference samples, which comprises:

[0037] The memory, the processor and the computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize the steps of the non-uniform low-illumination image enhancement method under zero reference samples provided by the present application.

[0038] The present application aims at solving the problem of low and unevenly distributed illumination and much noise in the acquired image caused by factors such as illumination environment and shooting device limitation, and proposes a non-uniform low-illumination image enhancement method under zero reference sample. In order to prevent noise from being amplified while enhancing the brightness of the image, a deep Retinex network is constructed by combining a robust Retinex model and deep learning, and the original image is decomposed into a reflection map, an illumination map and a noise map. By quantitatively describing the noise component contained in the image, the noise in the image is suppressed. In order to solve the problem of low and unevenly distributed brightness, an adaptive brightness mapping curve is proposed, and a brightness enhancement network is constructed to obtain the optimal mapping curve parameters according to the input illumination map, and different gray change rates are assigned to pixels with different brightness at different positions in the image, and an illumination map with enhanced brightness and naturalness is obtained, which is fused with the noise-free reflection map to obtain the final enhancement result. Since no pair of high-quality images is required and the brightness of the image can be adaptively adjusted according to different illumination environments, the present application can obtain an enhanced image with improved contrast and brightness.

[0039] The present application proposes a non-uniform low-illumination image enhancement method and system under zero reference sample, establishes an image decomposition network based on a robust Retinex model and deep learning, decomposes the input image into an illumination map, a reflection map and a noise map according to the image decomposition network, constructs an adaptive brightness mapping function, obtains the optimal brightness mapping curve corresponding to the illumination map, and obtains the enhanced illumination map and the enhanced image by multiplying the enhanced illumination map and the reflection map pixel by pixel, solving the technical problem of poor non-uniform low-illumination image enhancement effect. Since no pair of high-quality images is required and the brightness of the image can be adaptively adjusted according to different illumination environments, the present application can obtain a natural, contrast and brightness enhanced image from a non-uniform low-illumination image, and has strong generalization ability and can adapt to different scenes.

[0040] The beneficial effects of the present application include:

[0041] (1) The image decomposition network is established by combining a robust Retinex model and deep learning, and the noise is suppressed in the decomposition process.

[0042] (2) A hybrid no-reference decomposition loss function is designed to guide the decomposition network to decompose the image into a reflection map, an illumination map and a noise map.

[0043] (3) An adaptive brightness mapping curve is constructed, and different parameters in the curve are used to control the amplitude of low-light enhancement and high-light suppression, which can simultaneously realize low-light enhancement and high-light suppression of the illumination map.

[0044] (4) An adaptive light enhancement network is established to obtain the parameters of the optimal adaptive brightness mapping curve according to different light maps. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A flowchart of the non-uniform low-light image enhancement method under zero reference samples of the second embodiment of the present application is shown.

[0046] Figure 2 A decomposition network structure diagram of the second embodiment of the present application is shown.

[0047] Figure 3 An adaptive brightness mapping curve of the second embodiment of the present application is shown.

[0048] Figure 4 A result diagram of the image enhancement method using the second embodiment of the present application is shown, in which (a) and (b) are respectively the original input image and the enhanced result diagram after enhancement.

[0049] Figure 5 A material surface image enhancement result diagram of the third embodiment of the present application is shown.

[0050] Figure 6 A structure block diagram of the non-uniform low-light image enhancement system under zero reference samples of the embodiment of the present application is shown.

[0051] Reference signs:

[0052] 10, memory; 20, processor. DETAILED DESCRIPTION

[0053] In order to facilitate the understanding of the present application, the present application will be described more fully and completely below in conjunction with the drawings and preferred embodiments, but the scope of protection of the present application is not limited to the following specific embodiments.

[0054] The embodiments of the present application will be described in detail below in conjunction with the drawings, but the present application can be implemented in various different ways limited and covered by the claims.

[0055] Embodiment one

[0056] The non-uniform low-light image enhancement method under zero reference samples provided by the first embodiment of the present application comprises:

[0057] Step S101, an image decomposition network is established based on a robust Retinex model and deep learning.

[0058] Step S102, according to the image decomposition network, the input image is decomposed into a light map, a reflection map and a noise map.

[0059] Step S103, constructing an adaptive brightness mapping function to obtain an optimal brightness mapping curve corresponding to the illumination map.

[0060] Step S104, using the optimal brightness mapping curve corresponding to the illumination map to obtain an enhanced illumination map.

[0061] Step S105, multiplying the enhanced illumination map and the reflection map pixel by pixel to obtain an enhanced image.

[0062] The non-uniform low-light image enhancement method under zero reference samples provided by the embodiment of the present application solves the technical problem of poor non-uniform low-light image enhancement effect by establishing an image decomposition network based on a robust Retinex model and deep learning, decomposing an input image into an illumination map, a reflection map and a noise map according to the image decomposition network, constructing an adaptive brightness mapping function, obtaining an optimal brightness mapping curve corresponding to the illumination map, using the optimal brightness mapping curve corresponding to the illumination map to obtain an enhanced illumination map, and multiplying the enhanced illumination map and the reflection map pixel by pixel to obtain an enhanced image. Since there is no need for paired high-quality images and the image brightness can be adaptively adjusted according to different lighting environments, the embodiment of the present application can obtain natural, contrast and brightness enhanced images according to non-uniform low-light images, and has strong generalization and can adapt to different scenes.

[0063] Specifically, to solve the problems of non-uniform, low-light and noise caused by factors such as lighting environment and collection device limitation, the present application provides a non-uniform low-light image enhancement method under zero reference samples. The method mainly includes two parts: image decomposition considering noise and adaptive illumination enhancement. In image decomposition, the decomposition of the robust Retinex model is optimized based on deep learning and a hybrid no-reference decomposition loss function, and the amplification of noise is suppressed by separating the noise map and the structure-smoothed reflection map; in illumination enhancement, an adaptive brightness mapping curve suitable for non-uniform low-light enhancement is proposed, and the parameters thereof are adaptively adjusted by an illumination enhancement network, which can simultaneously realize low-light enhancement and high-light suppression of non-uniform illumination images.

[0064] Embodiment two

[0065] The present application provides a solution to the problem of non-uniform low-light image enhancement caused by factors such as lighting environment, shooting device limitation, etc., and the overall idea is as shown in Figure 1 The steps are as follows:

[0066] (1) An image decomposition network is established based on a robust Retinex model and deep learning, and noise is suppressed during decomposition.

[0067] (2) Design a mixed no-reference decomposition loss function to guide the image decomposition network to decompose the input image into an illumination map, a reflection map and a noise map.

[0068] (3) Considering the non-uniform brightness distribution of the original image, an adaptive brightness mapping function is designed to control the adjustment amplitude of different brightness in the illumination map by adjusting the parameters of the function.

[0069] (4) In order to obtain the optimal brightness mapping curve according to different input images, an illumination enhancement network is established to adaptively adjust the parameters of the mapping curve according to the input illumination map.

[0070] (5) According to the adaptive brightness mapping curve, the brightness enhanced illumination map is obtained, and the final enhanced result is obtained by multiplying the reflection map pixel by pixel.

[0071] The specific implementation scheme is as follows:

[0072] (1) Based on the robust Retinex model and deep learning, an image decomposition network is established to suppress noise in the decomposition process.

[0073] The classical Retinex model is based on color constancy to divide the image into a reflection map and an illumination map, wherein the reflection map represents the high-frequency information in the image and is an inherent property of the image, which is independent of the lighting condition, and the illumination map represents the low-frequency information of the image, which determines the maximum dynamic range of the brightness of each pixel.

[0074]

[0075] Wherein, respectively represent the original non-uniform low-light image, the reflection map and the illumination map, represents pixel-by-pixel multiplication.

[0076] Although the classical Retinex model has achieved good results in image enhancement, it does not consider the noise problem in the image, which can easily lead to the simultaneous amplification of the noise in the image in the enhancement process. Considering that image denoising can be regarded as a process of separating a noise-free image from the original image, the embodiment of the present application adds a noise component on the basis of the classical Retinex model to obtain a robust Retinex model.

[0077]

[0078] Wherein R, I, N respectively represent the reflection map, the illumination map and the noise map under the robust Retinex model.

[0079] It is a typical ill-posed problem to decompose three different components from a single image, and most of the existing decomposition methods adopt the variational method, which can effectively decompose different components in the image, but needs to manually adjust parameters according to different images, and increases the complexity of the algorithm, and the optimization process is time-consuming, and the neural network has strong self-learning ability and self-adaptive advantage, therefore, the image decomposition network is established based on the robust Retinex model and deep learning. The decomposition network contains 8 convolutional layers, wherein the 2nd and 3rd convolutional layers are connected through a jump connection and connected with the 5th and 6th layers, the activation function of the middle layer adopts ReLU to avoid gradient disappearance and speed up the training speed. In the output layer of the network, the output channels of the network for estimating the reflection map and the noise map are 3, and the activation functions are Sigmoid and Tanh respectively, so that the reflection map and the noise map are limited in the range of [0, 1] and [-1, 1] respectively, and the output channel of the network for estimating the illumination map is 1, and the activation function adopts Sigmoid.

[0080] (2) Design a hybrid no-reference decomposition loss function to guide the image decomposition network to decompose the input image into an illumination map, a reflection map and a noise map.

[0081] In actual production and life, it is generally difficult to obtain a pair of high-quality images. In order to effectively decompose the original non-uniform low-light image under the condition of no reference, a hybrid no-reference decomposition loss function is designed through reconstruction loss, reflection estimation loss, illumination estimation loss and noise estimation loss.

[0082]

[0083] Wherein L dec ,L rec ,L ref ,L ill ,L n Respectively represent the hybrid no-reference decomposition loss, the reconstruction loss, the reflection estimation loss, the illumination estimation loss and the noise estimation loss, λ1, λ2, λ3 are constants set in advance, respectively representing the weights of the reflection estimation loss, the illumination estimation loss and the noise estimation loss.

[0084] Different loss functions are specifically represented as follows:

[0085]

[0086] Wherein ||·||1 represents L1 norm, ||·||1 represents F norm. c is in {r, g, b}, representing the RGB three channels of the image. H(·) represents histogram equalization, Indicates the gradient sum in the horizontal and vertical directions, and β represents a constant, which is 10.

[0087] Lrec : reconstruction loss, measures the difference between the original image and the image obtained by inverse operation of the three components through the robust Retinex model, aims to ensure the effectiveness of the decomposition, so that the decomposed result can restore the original image.

[0088] L rec : reflection estimation loss, aims to obtain a reflection map with high contrast, rich information and noise suppression. The first item of the formula is to ensure that the obtained reflection map contains enough information, and the second item considers that the noise in the image generally has small gradient, while the non-noise content has large gradient, and minimizing the second item can suppress the noise in the reflection map.

[0089] L ill : illumination estimation loss, the first item of the formula is based on the bright channel prior, and the illumination map is regarded as the maximum brightness channel in the image. The second item considers that a good illumination map needs to ensure that the overall image is smooth and can reflect the structure of the image. By adding a weight related to the reflection map to the second item, the illumination smoothness is weighted smaller in places where the gradient of the reflection map is large, and weighted larger in places where the gradient is small.

[0090] L n : noise estimation loss, aims to suppress the intensity of noise. By solving the F-norm of the dot product of the original image and the noise map, the information and features of the original image can be used to constrain the overall intensity of the noise.

[0091] In order to decompose the original image into a reflection map, an illumination map and a noise map, a hybrid referenceless loss function is minimized as the objective function.

[0092] f dec = argmin L = argmin (L rec + λ1L ref + λ2L ill + λ3L n ) (5)

[0093] where f dec represents the objective function of the decomposition network.

[0094] By optimizing the objective function, the denoised reflection map, illumination map and noise map are obtained, as shown in Figure 2 .

[0095] (3) Considering the uneven brightness distribution of the original image, an adaptive brightness mapping function is designed, and the parameters of the function are adjusted to control the amplitude of the brightness adjustment in the illumination map.

[0096] Considering that the human eye's perception of external brightness is nonlinearly changed, assuming B1 is the illumination map of the original input image, and B2 is the illumination map of the well-exposed image, the mapping relationship between the two images can be expressed as:

[0097] B2 = g(B1) (6)

[0098] Where g(·) represents the brightness mapping curve, as shown in Figure 3 .

[0099] In order to realize the natural and reasonable mapping of brightness, the embodiment of the application gives three conditions that the brightness mapping curve of the non-uniform image needs to meet: 1) the function is monotonically increasing, and the difference in brightness order of the output image is consistent with the difference in brightness order of the input image; 2) the function can enhance the brightness of the low-light area, while suppressing the brightness of the high-light area, and the value range can be normalized to [0, 1] to prevent information overflow; 3) the function is derivable, and for pixels in the low-light area, the gradient decreases with the increase of brightness, and for pixels in the high-light area, the gradient increases with the increase of brightness.

[0100] In order to adapt to the non-uniformity of image brightness and meet the above conditions, an adaptive brightness mapping curve is designed based on the hyperbolic tangent function:

[0101]

[0102]

[0103] Where tanh(I) represents the hyperbolic tangent function, the value range is [-1, 1], ω controls the weight of low-light enhancement, and k1, k2 can simulate exposure, which are used to adjust the amplitude of low-light enhancement and high-light suppression, respectively.

[0104] (4) In order to obtain the best brightness mapping curve according to different input images, an illumination enhancement network is established to adaptively adjust the parameters of the mapping curve according to the input illumination map.

[0105] In order to adaptively adjust the parameters of the brightness mapping curve according to different illumination maps, a lightweight illumination enhancement network is constructed. The network structure is simple, including 7 convolutional layers and a fully connected layer, the convolutional layers are connected through ReLU activation function and max pooling layer, and the fully connected layer outputs the best parameters ω, k1, k2. Combined with the three conditions of the brightness mapping curve, the illumination enhancement loss function for network training is given.

[0106] L enc = L E + η1L N + η2L S (9)

[0107] where L E , N , S respectively represent the exposure control loss, naturalness loss and structure preserving illumination smoothness loss, and η1,η2 respectively represent the weights of different loss functions, which are constants in the experiment.

[0108] In order to improve the brightness of the image and maintain the brightness value near the good exposure, the brightness map is divided into K 4x4 non-overlapping image blocks, and an exposure loss is designed.

[0109]

[0110] where sign(·) represents the sign function, which takes 1 when greater than 0, -1 when less than 0, and 0 when 0. E represents the good exposure, which can be taken as [0.5, 0.7]. O i and I i respectively represent the brightness value of the i-th image block in the enhanced and original illumination map.

[0111] In order to maintain the consistency of the brightness order difference between the enhanced image and the original image, the naturalness loss function is given by measuring the brightness difference between the four-neighbor image block centered at pixel i and the image block i.

[0112]

[0113] where Ω(i) represents the four-neighbor spatial space centered at pixel i, O j ,I j respectively represent the brightness value of the four-neighbor spatial space of the enhanced and original illumination map.

[0114] In addition, the enhanced brightness map still needs to satisfy the overall smoothness and be able to reflect the structure of the image, so the second term of the illumination estimation loss in the decomposition remains consistent, and the structure preserving illumination smoothness loss is given.

[0115]

[0116] where represents the gradient sum in the horizontal and vertical directions of the enhanced illumination map.

[0117] By minimizing the illumination enhancement loss function, the objective function of the brightness enhancement network is obtained.

[0118] f enh = argmin(L E + η1L N + η2L S ) (13)

[0119] where fdec Objective function representing the brightness enhancement network.

[0120] (5) According to the adaptive brightness mapping curve, the light map after brightness enhancement is obtained, and the final enhancement result is obtained by multiplying the reflection map pixel by pixel.

[0121] After obtaining the denoised reflection map and the enhanced brightness map, the enhanced result is obtained according to the inverse operation of the Retinex model.

[0122]

[0123] Wherein Indicates the enhanced image. Figure 4 The figure shows the result of non-uniform low-light image enhancement using the method of the embodiment. Wherein Figure 4 (a) and (b) in (a) and (b) respectively represent the original input image and the enhanced result of the enhanced image.

[0124] Example three

[0125] This embodiment takes a 2650m 3 blast furnace of an iron mill as an experimental platform, and collects the material surface images of the blast furnace smelting process from the field. The brightness and contrast of the material surface image are improved by using the non-uniform low-light image enhancement method proposed in the application. The specific implementation steps are as follows:

[0126] 1. Establish a robust Retinex model for the blast furnace material surface image, and set the parameters of the image decomposition network, such as the number of iterations;

[0127] 2. By minimizing the hybrid no-reference decomposition loss function, the reflection map, the light map and the noise map under different training iteration numbers are obtained;

[0128] 3. Continuously update the reflection map, the light map and the noise map until the number of iterations is reached or the loss function is minimized;

[0129] 4. Input the original light map to the brightness enhancement network, and set the network parameters, such as the number of iterations;

[0130] 5. By minimizing the loss function of the light enhancement network, continuously optimize the brightness mapping process until the set number of iterations is reached or the loss function is minimized, and the best adaptive brightness mapping curve parameters are obtained;

[0131] 6. Multiply the noise-free reflection map and the enhanced brightness map pixel by pixel to obtain the final enhancement structure.

[0132] Figure 5A result diagram of image enhancement of the non-uniform low-lighting material surface image by using the method of the embodiment is shown. It should be noted that the blast furnace material surface image enhancement provided by the embodiment is only an industrial implementation case given by the present application, and is not limited to the blast furnace material surface. Since the embodiment of the present application does not need a pair of high-quality images as a reference, and can adaptively adjust the image brightness according to different lighting environments, the contrast and brightness of different lighting environments and application scenarios can be improved, and the generalization is good.

[0133] With reference to Figure 6 The non-uniform low-lighting image enhancement system under zero reference samples provided by the embodiment of the present application includes a memory 10, a processor 20, and a computer program stored in the memory 10 and executable on the processor 20. When the processor 20 executes the computer program, the steps of the non-uniform low-lighting image enhancement method under zero reference samples provided by the embodiment are implemented.

[0134] The specific working process and working principle of the non-uniform low-lighting image enhancement system under zero reference samples of the embodiment can refer to the working process and working principle of the non-uniform low-lighting image enhancement method under zero reference samples of the embodiment.

[0135] The above is only the preferred embodiment of the present application, and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A non-uniform low-light image enhancement method under zero reference samples, characterized in that, The method comprises: An image decomposition network is established based on a robust Retinex model and deep learning; An input image is decomposed into an illumination map, a reflection map and a noise map according to the image decomposition network; An adaptive brightness mapping function is constructed to obtain an optimal brightness mapping curve corresponding to the illumination map, wherein a specific formula of the adaptive brightness mapping function is: wherein g(·) represents a brightness mapping curve, I represents an illumination map, tanh(I) represents a hyperbolic tangent function, the value range is [-1, 1], ω controls the weight of low-light enhancement, k1 and k2 are respectively used for adjusting the amplitudes of low-light enhancement and high-light suppression, and the adaptive brightness mapping function is constructed to obtain the optimal brightness mapping curve corresponding to the illumination map, which comprises: An illumination enhancement network is established according to the adaptive brightness mapping function, the illumination enhancement network comprises seven convolutional layers and one fully connected layer, the convolutional layers are connected through ReLU activation functions and maximum pooling layers, and the fully connected layer outputs optimal parameters ω, k1 and k2; The illumination enhancement network is trained, and the parameters of the adaptive brightness mapping function are adjusted according to the trained illumination enhancement network, so as to obtain the optimal brightness mapping curve corresponding to the illumination map; The optimal brightness mapping curve corresponding to the illumination map is used to obtain an enhanced illumination map; The enhanced illumination map and the reflection map are multiplied pixel by pixel to obtain an enhanced image.

2. The non-uniform low-light image enhancement method under zero reference samples according to claim 1, characterized in that, The robust Retinex model is specifically: Wherein, S represents the input image, R, I, N respectively represent the reflection map, the illumination map and the noise map under the robust Retinex model, represents pixel-by-pixel multiplication.

3. The non-uniform low-light image enhancement method under zero reference samples according to claim 2, characterized in that, A specific formula of a loss function of the robust Retinex model is: f dec = argminL dec = argmin(L rec + λ1L ref + λ2L ill + λ3L n ), Among them, f dec L represents the loss function of a robust Retinex model. dec ,L rec ,L ref ,L ill ,L n Let λi represent the hybrid no-reference decomposition loss, reconstruction loss, reflection estimation loss, illumination estimation loss, and noise estimation loss, respectively. Let λ1, λ2, and λ3 be the preset weights for reflection estimation loss, illumination estimation loss, and noise estimation loss, respectively. Let S represent the input image, and R, I, and N represent the reflection map, illumination map, and noise map under the robust Retinex model, respectively. Let ||·||1 represent the L1 norm. F Let F represent the norm, c∈{r,g,b}, represent the RGB channels of the image, and H(·) represent histogram equalization. This represents the sum of the gradients in the horizontal and vertical directions, where β represents a user-defined constant.

4. The non-uniform low-light image enhancement method under zero reference samples of claim 3, characterized in that, A specific formula of an objective function for training the illumination enhancement network is: f enh = argmin(L E + η1L N + η2L S ), Among them, f dec Let L represent the objective function of the brightness enhancement network. E ,L N ,L S Let η1 and η2 represent the exposure control loss, naturalness loss, and illumination smoothness loss, respectively. Let η1 and η2 represent the weights of the naturalness loss function and the illumination smoothness loss function, respectively. Let sign(·) represent the sign function, which takes the value 1 when it is greater than 0, -1 when it is less than 0, and 0 when it is 0. Let E represent the exposure value, and O represent the illumination smoothness loss value. i and I i Let Ω(i) represent the brightness value of the i-th image patch in the enhanced and original illumination maps, respectively, and let O(i) represent the four-neighbor space centered at pixel i. j ,I j These represent the brightness values ​​of the four neighborhood spaces of the enhanced and original lighting maps, respectively. This represents the sum of the horizontal and vertical gradients of the enhanced lighting map, where β represents a custom constant.

5. The non-uniform low-light image enhancement method under zero reference samples according to claim 4, characterized in that, A specific formula for multiplying the enhanced illumination map and the reflection map pixel by pixel to obtain the enhanced image is: wherein denotes the enhanced image, denotes the enhanced illumination map, and R denotes the reflectance map under the robust Retinex model.

6. A non-uniform low-light image enhancement system under zero reference samples, the system comprising: A memory (10), a processor (20), and a computer program stored on the memory (10) and executable on the processor (20), wherein the processor (20) implements the steps of the method of any one of claims 1 to 5 when executing the computer program.

Images