Non-uniform low-illumination image enhancement method and system

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a non-uniform low-illumination image enhancement method and system. The method comprises the following steps: step 1, constructing a network model: constructing a non-uniform low-illumination image enhancement network model, wherein the model comprises a Retinex network, an illumination image enhancement network, a reflection image enhancement network and a multi-scale enhancement network; step 2, preparing a data set: dividing a training set and a test set by using an LOL-v1 data set, and preprocessing the training set and the test set; and performing fine adjustment on the model by the LOL-v2 data set. According to the invention, based on deep learning, an efficient two-way feature enhancement network is designed and is used for improving the performance of image features. On the basis of Retinex network decomposition, an input image is decomposed into an illumination image component and a reflection image component, and the characteristics and channel dimensions of the illumination image component and the reflection image component are modulated respectively, so that the characteristics of a low-illumination image are effectively enhanced. According to the method, the quality and definition of the image can be remarkably improved, and illumination and reflection details are brighter.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to a method and system for enhancing an uneven low-illuminance image. Background Art

[0002] Uneven low-light image enhancement has always been a hot research topic in the field of computer vision. It aims to reconstruct high-quality images with rich edge, texture and other detailed features from a single or multiple uneven low-light images through a series of image processing and deep learning techniques. Since the deep convolutional neural network was applied to the low-light image enhancement algorithm, its performance has been greatly improved compared with the traditional low-light image enhancement algorithm based on Retinex theory and histogram specification. However, the current low-light image enhancement algorithm still needs to be improved and optimized in terms of actual scene application, algorithm performance, model quality evaluation standards, etc.

[0003] The Chinese patent publication number is "CN114897753B", and the name is "A low-illumination image enhancement method". The low-illumination color image A is converted into HSV space, the brightness component V is extracted, and it is converted into a grayscale image A-1. A-1 is enhanced by homomorphic filtering to obtain A-2, and its brightness V-1 is extracted. Taking A-2 as the guide image, guided filtering is performed to enhance the brightness V to obtain V-2. Subsequently, Retinex enhancement is performed on V-2 to obtain V-3. Finally, the grayscale images of V-1, V-2 and V-3 are PCA fused to obtain the fused brightness V-4. Convert back to RGB space to obtain an enhanced color image. This method only enhances the brightness component in the HSV space during image enhancement. The image enhancement method adopted is too simple and superficial, and does not take into account the influence of a large number of uncontrollable factors in the real environment on the image, resulting in unsatisfactory quality of the image after enhancement in an environment affected by uneven low-illumination scenes. Therefore, we propose a method for uneven low-illumination image enhancement to solve the above problems. Summary of the invention

[0004] 1. Technical issues to be solved

[0005] In view of the shortcomings of the prior art, the present invention provides a method and system for enhancing uneven low-illuminance images. The method decomposes the uneven low-illuminance image into an illumination image and a reflection image, and then enhances the image quality of the two images through different image enhancement networks, so that the enhanced uneven low-illuminance image has better clarity and better quality, is more in line with human visual observation, and solves the problems raised in the above-mentioned background technology.

[0006] (II) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0008] A method for enhancing an uneven low-light image comprises the following steps:

[0009] Step 1, building a network model: building an uneven low-illumination image enhancement network model, which includes a Retinex network, an illumination image enhancement network, a reflection image enhancement network, and a multi-scale enhancement network;

[0010] Step 2, prepare the data set: use the LOL-v1 data set as the first data set, divide it into a training set and a test set; use the LOL-v1 data set as the second data set to fine-tune the model;

[0011] Step 3, training the network model: the data set prepared in step 2 is passed through the Retinex network and input into the network model built in step 1 for training;

[0012] Step 4, fine-tuning the model: retraining and fine-tuning the network model using the second data set to obtain the final model;

[0013] Step 5, save the model: solidify the parameters of the final model and save the model.

[0014] Furthermore, in step 1, the illumination image enhancement network uses pooling layers, convolution layers, activation functions, fully connected layers, channel weight layers, deblurring modules and residual blocks to extract a large number of features from the illumination image components obtained after decomposition by the Retinex network; the reflection image enhancement network decomposes the reflection image components obtained after decomposition by the Retinex network into three image components of different frequencies through discrete wavelet transform, and uses convolution layers, activation functions and global average pooling layers to extract a large number of features; the multi-scale enhancement network is composed of pooling layers, convolution layers and activation functions, and is used to perform high-quality enhancement on the enhanced reflection image.

[0015] Furthermore, in the step 1, the illumination image enhancement network is composed of an average pooling layer, a maximum pooling layer, a fully connected layer one, a fully connected layer two, a channel weight layer, a convolutional layer one, a convolutional layer two, a residual block one, a residual block two, a residual block three, a residual block four, a deblurring module and an R-type function and an S-type function; the reflection image enhancement network is composed of a discrete wavelet transform module, an inverse wavelet transform module, a convolutional layer three, a convolutional layer four, a convolutional layer five, a convolutional layer six, a convolutional layer seven, a global average pooling layer and an R-type function and an S-type function; the multi-scale enhancement network is composed of a maximum pooling layer, a convolutional layer eight, a convolutional layer nine, a convolutional layer ten, a convolutional layer eleven, a convolutional layer twelve, a convolutional layer thirteen, a convolutional layer fourteen, a convolutional layer fifteen, a convolutional layer sixteen, a convolutional layer seventeen, a convolutional layer eighteen, a convolutional layer nineteen and an L-type function, a T-type function, and an M-type function.

[0016] Furthermore, in step 2, the prepared data set is processed, and the first data set is divided into a training set and a test set as network input.

[0017] Furthermore, in step 3, the loss function is selected to use a composite loss function, the illumination image enhancement network uses a smoothness loss function, the reflection image enhancement network uses a structural similarity loss function, and the multi-scale enhancement network uses a mean square error loss function; the choice of loss function affects the quality of the model, can truly reflect the difference between the predicted value and the true value, and can correctly feedback the quality of the model. The training network model also includes evaluating the quality of the uneven low-light image enhancement results and the degree of image distortion through evaluation indicators to measure the role of the uneven low-light image enhancement network.

[0018] Furthermore, in step 4, the first data set is used to train the network to enhance the robustness of the network.

[0019] Furthermore, the quality of the algorithm enhancement results and the degree of image distortion are evaluated by evaluation indicators.

[0020] A non-uniform low-illumination image enhancement system, the system comprising:

[0021] Visual perception module: used to load the dataset, and the loaded item is uneven low-light images;

[0022] Intelligent processing module: used to process the loaded uneven low-light images and adjust the size of each image. The image processing module crops the image to a size of 256×256 and fixes the size of the input image;

[0023] Image optimization module: used to train uneven low-light images after decomposition by the Retinex network, including illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network, and obtain the final enhancement network model through iteration;

[0024] Image output module: used to display the enhanced high-quality image and output the enhanced high-quality image using electronic equipment.

[0025] The visual perception module is the beginning of the whole process. It is responsible for loading the data sets, which contain the uneven low-light images to be processed. The goal of this module is to collect and prepare image data for subsequent processing.

[0026] The intelligent processing module, once the image is loaded, is responsible for preprocessing the uneven low-light image. The image size is adjusted to 256×256 pixels, which ensures the consistency and standardization of the input image. This standard size provides a consistent benchmark for subsequent processing.

[0027] The image optimization module, which is the core of the whole process, includes three main networks: illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network. These networks process loaded and pre-processed images through iterative training; the illumination image enhancement network is responsible for enhancing the illumination image obtained after the Retinex network decomposition, while the reflection image enhancement network is dedicated to enhancing the reflection image obtained after the Retinex network decomposition. The multi-scale enhancement network is responsible for further improving the quality of the image processed by the reflection image enhancement network. The goal of this module is to generate an optimized uneven low-light image enhancement model;

[0028] The image output module is responsible for displaying the enhanced high-quality image. This module uses electronic equipment to output the processed high-quality image so that users can observe and evaluate the final result. These four modules cooperate with each other to form a process, which ultimately realizes the enhanced processing of uneven low-light images through preprocessing, training and output.

[0029] The above system is stored in a computer storage medium, and the computer-readable storage medium stores computer executable program code. When the code is run, the steps of the non-uniform low-illumination image enhancement method are implemented.

[0030] The above-mentioned storage medium is located in a computer device, which includes an input / output unit, a central processing unit, a memory and a communication component, etc. A computer program is stored in the memory. When the computer program is executed by the processor, the steps of the non-uniform low-illumination image enhancement method are implemented.

[0031] (III) Beneficial effects

[0032] Compared with the prior art, the present invention provides a method and system for enhancing uneven low-illumination images, which have the following beneficial effects:

[0033] 1. Based on deep learning, this paper designs an efficient two-way feature enhancement network to improve the performance of image features. Based on the decomposition of the Retinex network, we decompose the input image into an illumination image component and a reflection image component, and modulate their features and channel dimensions respectively, thereby effectively enhancing the features of low-light images. This method can significantly improve the quality and clarity of the image, making the illumination and reflection details more distinct.

[0034] 2. The present invention helps to reduce network parameters by adding dense residuals and jump connections, making the depth of the network shallower and the number of network parameters smaller. It also designs a multi-scale enhancement network and a deblurring module, which ultimately makes the entire network structure simple and improves the quality of the enhanced image.

[0035] 3. The present invention designs a composite loss function composed of mean square error loss, smoothness loss and structural similarity loss, which optimizes the image quality from three aspects: reducing pixel differences, structural errors and visual perception, while making the image more realistic. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the process structure of the method of the present invention;

[0037] Figure 2 It is a schematic diagram of the structure of the uneven low-illumination image enhancement model of the present invention;

[0038] Figure 3 This is a schematic diagram of the illumination image enhancement network structure of the present invention;

[0039] Figure 4 This is a schematic diagram of the reflected image enhancement network structure of the present invention;

[0040] Figure 5 This is a schematic diagram of the multi-scale enhancement network structure of the present invention;

[0041] Figure 6 This is a schematic diagram of the network structure of the defuzzification module of the present invention;

[0042] Figure 7 Schematic diagram of the residual block network structure of the present invention;

[0043] Figure 8 A schematic diagram of the structure of the image enhancement system provided by the present invention;

[0044] Fig. 9 A schematic diagram of the structure of a device for performing image enhancement tasks provided by the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Example 1

[0047] like Figure 1As shown, the embodiment 1 of the present invention provides a method for enhancing an uneven low-light image, which specifically includes the following steps:

[0048] Step 1: Build network models: illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network;

[0049] like Figure 2 , provides a schematic diagram of the uneven low-light image enhancement model structure, where Figure 3 , provides a schematic diagram of the illumination image enhancement network structure.

[0050] The illumination image enhancement network consists of an average pooling layer, a maximum pooling layer, a fully connected layer 1, a fully connected layer 2, a channel weight layer, a convolutional layer 1, a convolutional layer 2, a residual block 1, a residual block 2, a residual block 3, a residual block 4, a deblurring module, and an R-type function and an S-type function. The channel weight layer uses a one-dimensional convolution kernel, which is equivalent to sampling the original input in the channel dimension; the internal channel number of the fully connected layer 1 is set to 256, and the activation function is the R-type function; the internal channel number of the fully connected layer 2 is set to 512, and the activation function is the S-type function; the convolution kernel size of the convolutional layer 1 is 3×3, the step size is 1, and the activation function is the R-type function; the convolution kernel size of the convolutional layer 2 is 3×3, the step size is 1, and the activation function is the S-type function; the four residual blocks are all composed of convolutional layers and T-type functions, such as Figure 7 , provides a residual block network structure; such as Figure 6 , provides a schematic diagram of the network structure of the deblurring module, which consists of two 1×1 convolutional layers, an average pooling layer, a maximum pooling layer, six convolutional layers, two S-type functions, two T-type functions, and three batch normalization layers;

[0051] like Figure 4 , provides a schematic diagram of the structure of the reflection image enhancement network, which consists of a discrete wavelet transform module, an inverse wavelet transform module, three sub-band blocks, five convolution layers, and a global average pooling layer; the convolution kernel size of the convolution layer three is 3×3, the step size is 1, and the activation function is an R-type function, the convolution kernel size of the convolution layer four is 3×3, the step size is 1; the convolution kernel size of the convolution layer five is 3×3, the step size is 1, the convolution kernel size of the convolution layer six is ​​3×3, the step size is 1; the convolution kernel size of the convolution layer seven is 3×3, the step size is 1, and the activation function is an S-type function; the discrete wavelet transform module converts the input to the frequency domain and divides it into three sub-bands, enhances them respectively, and then transforms them to the time domain through the inverse wavelet transform module;

[0052] like Figure 5, a schematic diagram of the multi-scale enhancement network structure is provided. The multi-scale enhancement network consists of a maximum pooling layer and a large number of convolutional layers. The convolution kernel size of convolutional layer eight is 3×3, the step size is 1, and the activation function is an L-type function; the convolution kernel size of convolutional layer nine is 3×3, the step size is 1, and the activation function is an M-type function; the convolution kernel size of convolutional layer ten is 3×3, the step size is 1, and the activation function is a T-type function; the convolution kernel size of convolutional layer eleven, convolutional layer twelve, convolutional layer thirteen, convolutional layer fourteen, convolutional layer fifteen, convolutional layer sixteen, convolutional layer seventeen, convolutional layer eighteen, and convolutional layer nineteen are all 3×3, with a step size of 1.

[0053] Among them, the R-type function calculation expression is f(x) ReLU =max(0,x); the S-type function calculation expression is The calculation expression of L-type function is: The calculation expression of T-type function is: The calculation expression of M-type function is f(x) Mish =x*tanh(ln(1+e x )).

[0054] Step 2: Prepare the data set: Train the network with the first data set. When using the LOL-v1 data set as the first data set, divide it into a training set and a test set.

[0055] Step 3, training network model: train the uneven low-light image enhancement model, preprocess the data set prepared in step 2, adjust the size of each image in the data set, fix the size of the input image, and input the processed data set into the network model built in step 1 for training.

[0056] During the training process, the loss function value to be minimized and the optimal evaluation index are selected. Specifically, the model parameters are considered to have been pre-trained by minimizing the loss function of the network output image and the label until the number of training times reaches the set threshold or the value of the loss function reaches the set range, and the model parameters are saved. At the same time, the optimal evaluation index is selected to measure the accuracy of the algorithm and evaluate the performance of the system; during the training process, the loss function is selected to use a composite loss function. The illumination image enhancement network uses a smoothness loss function; the reflection image enhancement network uses a structural similarity loss function; and the multi-scale enhancement network uses a mean square error loss function.

[0057] In the illumination image enhancement network, self-supervision is performed by the smoothness loss function. It is generally believed that the changes in natural images are smooth, and its working principle is to minimize the gradient between pixels and neighborhoods:

[0058]

[0059] Where N is the number of iterations; and Represent the gradient values ​​in the horizontal and vertical directions respectively, and out_img represents the output image.

[0060] In the reflection image enhancement network, self-supervision is performed by the structural similarity loss function. It is also a frequently used loss function, which generally measures the difference between the GT image and the output image. The structural similarity loss evaluates the image quality by integrating brightness, contrast and structure, while the structural similarity loss removes the brightness dimension, which can help to preserve the structure and texture well during the model training process:

[0061]

[0062] Among them, M: represents different scales; μ p ,μ g Represent the mean of the output image and GT respectively; σ p ,σ g Represent the standard deviation of the output image and GT respectively; σ pg represents the covariance between the two; β m and γ m , c1 and c2 are both constants.

[0063] In the multi-scale enhancement network, the mean square error loss function is used:

[0064]

[0065] Among them, y i is the real output data, It is the output data of the model fitting. From here we can see that the closer the loss is to 0, the better the model and fit are, and the more successful the data prediction is.

[0066] Therefore, the loss function is defined as:

[0067] L=0.4*L Smoothness-Loss +0.3*L SSIM +0.3*L MSE

[0068] The choice of loss function affects the quality of the model. It can truly reflect the difference between the predicted value and the true value, and can correctly feedback the quality of the model.

[0069] The peak signal-to-noise ratio and structural similarity are suitable evaluation indicators, which can effectively evaluate the quality of the uneven low-illumination image enhancement results and the degree of image distortion, and measure the role of the uneven low-illumination image enhancement network.

[0070] The peak signal-to-noise ratio is based on the error between corresponding pixels, that is, an error-sensitive image quality evaluation; structural similarity measures image similarity from three aspects: brightness, contrast, and structure. It is an indicator used to measure the similarity between two digital images.

[0071] The peak signal-to-noise ratio and structural similarity are defined as follows:

[0072]

[0073] Among them, MSE is the mean square error between the original image and the processed image, MAX I Indicates the maximum value of the image color, which is 255 for an 8-bit sampling point.

[0074]

[0075] Among them, μ x , μ y Represent the mean of images x and y respectively, and Represents the variance of images x and y, σ xy represents the covariance of images x and y, C1 and C2 are constants.

[0076] The number of training times is set to 400, the learning rate of the first 200 training processes is set to 0.0001, and the learning rate of the next 200 training processes is gradually reduced from 0.0001 to 0; the upper limit of the number of images input to the network each time is mainly determined by the performance of the computer graphics processor. Generally, the number of images input to the network each time is in the range of 8-16, which can make the network training more stable and the training results better, and can ensure the rapid fitting of the network. The Adam optimizer is selected as the network parameter optimizer. Its advantages are simple implementation, efficient calculation, low memory requirements, and parameter updates are not affected by the scaling transformation of the gradient, making the parameters relatively stable.

[0077] Step 4, fine-tuning the model: Use the second dataset to train and fine-tune the model to obtain stable and usable model parameters, further improve the model's uneven low-light image enhancement capabilities; ultimately make the model-enhanced image quality better; use the LOL-v2 dataset in the process of fine-tuning the model parameters.

[0078] Step 5: Save the model: solidify the finalized model parameters.

[0079] If performing uneven low-light image enhancement operations, the final high-quality image can be obtained by directly inputting the image into the network.

[0080] Example 2

[0081] like Figure 8As shown, an embodiment of the present invention provides a non-uniform low-illumination image enhancement system, comprising:

[0082] Visual perception module: used to load the data set, and the loaded items are the uneven low-light images to be preprocessed;

[0083] Intelligent processing module: used to pre-process the loaded uneven low-light images and adjust the size of each image. The image processing module crops the image to a size of 128×128 and fixes the size of the input image;

[0084] Image optimization module: used to train the loaded and processed uneven low-light images, including illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network, and obtain the final image enhancement model through iteration;

[0085] Image output module: used to display the enhanced high-quality image and output the enhanced high-quality image using electronic equipment.

[0086] The visual perception module is the beginning of the whole process. It is responsible for loading the data set, which contains the uneven low-light images to be processed. The goal of this module is to collect and prepare image data for subsequent processing.

[0087] Next is the intelligent processing module. Once the image is loaded, the task of this module is to pre-process the uneven low-light image. The image size is adjusted to 128×128 pixels, which ensures the consistency and standardization of the input image. Through this standard size, a consistent benchmark is provided for subsequent processing;

[0088] The third is the image optimization module, which is the core of the whole process. It includes three main networks: illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network. These networks process and load pre-processed images through iterative training; the illumination image enhancement network is responsible for enhancing the illumination image obtained by Retinex network decomposition, while the reflection image enhancement network is dedicated to enhancing the reflection image obtained by Retinex network decomposition. The multi-scale enhancement network is mainly used to further enhance the quality of the output after passing through the reflection image enhancement network. The goal of this module is to generate a two-way image enhancement model;

[0089] Finally, there is the image output module, which is responsible for displaying the enhanced high-quality image. This module uses electronic devices to output the processed high-quality image so that users can observe and evaluate the final result.

[0090] These four modules work together to form a process, which ultimately realizes the enhanced processing of uneven low-light images through preprocessing, training and output.

[0091] Example 3

[0092] An embodiment of the present invention provides a computer storage medium, which is a variety of media for storing data, programs and files. According to their characteristics, uses and working methods, these media include:

[0093] Hard disk drives (HDD), solid-state drives (SSD), flash drives (USB), optical disks, memory (RAM), and cloud storage.

[0094] HDD is a mechanical storage device with large capacity and low price, but slow speed, and is often used for long-term data storage;

[0095] SSD uses flash memory chips to store data, has no moving parts, has faster read and write speeds, and is more durable, but is more expensive and is widely used in laptops, desktops, and servers.

[0096] Flash drives are portable devices that use flash memory technology and are suitable for temporary data storage and transfer;

[0097] Optical discs include CDs, DVDs, and Blu-rays. They use lasers to read data and are suitable for long-term storage of data, audio, and video. Although they have a smaller capacity, they are still useful in some scenarios.

[0098] Memory (RAM) is a temporary storage medium of the computer, used to store currently running programs and data. It has a fast speed, but the data will be lost after power failure;

[0099] Cloud storage stores data on a remote server and allows access at any time through the Internet, providing convenient data management and backup functions. The computer storage medium may store one or more programs, which, when executed, can implement the uneven low-light image enhancement method in the embodiment of the present invention.

[0100] Example 4

[0101] like Fig. 9As shown, an electronic device proposed in one embodiment of the present invention represents a structural schematic diagram of an electronic device suitable for implementing an embodiment of the present application. Computer equipment refers to an electronic device system for processing data, executing programs and completing various tasks, including multiple components for inputting, processing, storing and outputting data. The electronic device includes a central processing unit, which can perform various appropriate actions and processes such as steps one to five of the technical solution according to the program for performing the uneven low-light image enhancement stored in the memory. Communication components play an important role in connection and data transmission in computer systems, and they enable different devices, systems or networks to exchange data and transfer information. The power supply component provides power to supply various hardware devices to operate. The sensor component plays an important role in collecting, monitoring and converting real-world information into digital signals in the computer system. The display component is an important component in a computer system for outputting images, text and other visual information.

[0102] The present invention constructs a method and system for enhancing uneven low-light images, which can directly generate high-quality images from uneven low-light images without going through other intermediate steps. Under the same conditions, the feasibility and superiority of the method are further verified by calculating the relevant indicators of the image obtained by the existing method. The comparison of the relevant indicators of the existing technology and the method proposed in the present invention is shown in the following table;

[0103] Comparison of relevant indicators between the prior art and the method proposed in the present invention

[0104]

[0105] It can be seen from the table that the method proposed in the present invention has higher peak signal-to-noise ratio and structural similarity than the existing method. These indicators also further illustrate that the method proposed in the present invention has better uneven low-illumination image enhancement quality.

[0106] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for enhancing uneven low-illumination images, characterized in that: The following steps are included: Step 1, building a network model: building an uneven low-illumination image enhancement network model, which includes a Retinex network, an illumination image enhancement network, a reflection image enhancement network, and a multi-scale enhancement network; Step 2, prepare the data set: use LOL-v1 as the first uneven low-light image data set, divide it into a training set and a test set, and preprocess it; LOL-v2 is used as the second uneven low-light image dataset to fine-tune the model; Step 3, training the network model: input the data set prepared in step 2 into the network model constructed in step 1 for training until the number of training times reaches the initial set threshold or the value of the loss function reaches the preset range, the network model is considered to have been trained, and the network model parameters are saved; Step 4, fine-tuning the model: retraining and fine-tuning the network model using the second uneven low-illumination image data set, optimizing the network model parameters, improving the image enhancement performance, and obtaining a network model that can achieve the optimal effect; Step 5, save the model: solidify the final model parameters and save the model. After the model inputs an uneven low-light image, it can output a corresponding high-quality image.

2. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In step 1, the illumination image enhancement network extracts a large number of features from the illumination image components obtained after decomposition by the Retinex network using a pooling layer, a convolutional layer, an activation function, a fully connected layer, a channel weight layer, a deblurring module, and a residual block; The reflected image enhancement network decomposes the reflected image component obtained after decomposition by the Retinex network into three image components of different frequencies through discrete wavelet transform, and is used to extract a large number of features by using convolutional layers, activation functions and global average pooling; The multi-scale enhancement network, composed of a pooling layer, a convolutional layer and an activation function, is used to perform high-quality enhancement on the enhanced reflected image.

3. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In the step 1, the illumination image enhancement network is composed of an average pooling layer, a maximum pooling layer, a fully connected layer 1, a fully connected layer 2, a channel weight layer, a convolutional layer 1, a convolutional layer 2, a residual block 1, a residual block 2, a residual block 3, a residual block 4, a deblurring module and an R-type function and an S-type function; the reflection image enhancement network is composed of a discrete wavelet transform module, an inverse wavelet transform module, a convolutional layer 3, a convolutional layer 4, a convolutional layer 5, a convolutional layer 6, a convolutional layer 7, a global average pooling layer and an R-type function and an S-type function; the multi-scale enhancement network is composed of a maximum pooling layer, a convolutional layer 8, a convolutional layer 9, a convolutional layer 10, a convolutional layer 11, a convolutional layer 12, a convolutional layer 13, a convolutional layer 14, a convolutional layer 15, a convolutional layer 16, a convolutional layer 17, a convolutional layer 18, a convolutional layer 19 and an L-type function, a T-type function and an M-type function.

4. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In step 2, the prepared data set is preprocessed, the size of each image in the data set is adjusted, and the size of the input image is fixed as the network input.

5. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In step 3, the loss function is selected to use a composite loss function, the illumination image enhancement network uses a smoothness loss function, the reflection image enhancement network uses a structural similarity loss function, and the multi-scale enhancement network uses a mean square error loss function; the choice of loss function affects the quality of the model, can truly reflect the difference between the predicted value and the true value, and can correctly feedback the quality of the model.

6. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In the step 3, the training network model also includes evaluating the quality of the uneven low-light image enhancement result and the degree of image distortion through evaluation indicators to measure the role of the enhancement network.

7. The method for enhancing uneven low-illumination images according to claim 1, characterized in that: In step 4, the LOL-v2 data set is used to train the network to enhance the robustness of the network.

8. A non-uniform low illumination image enhancement system, characterized in that: The system includes: Visual perception module: used to load the data set, and the loaded items are the uneven low-light images to be preprocessed; Intelligent processing module: used to pre-process the loaded uneven low-light images, adjust the size of each image, and fix the size of the input image; Image optimization module: used to train the loaded and processed uneven low-light images and obtain the final image enhancement model through iteration; Image output module: used to display the enhanced high-quality image and output the enhanced high-quality image using electronic equipment; The visual perception module is the beginning of the whole process. It is responsible for loading the data set, which contains the uneven low-light images to be processed. The goal of this module is to collect and prepare image data for subsequent processing. Next is the intelligent processing module. Once the image is loaded, the task of this module is to pre-process the uneven low-light image. The image size is adjusted to 256×256 pixels, which ensures the consistency and standardization of the input image. Through this standard size, a consistent benchmark is provided for subsequent processing; The third is the image optimization module, which is the core of the whole process. It includes three main networks: illumination image enhancement network, reflection image enhancement network and multi-scale enhancement network. These networks process loaded and pre-processed images through iterative training; the illumination image enhancement network is responsible for enhancing the illumination image obtained after decomposition by the Retinex network, while the reflection image enhancement network is dedicated to enhancing the reflection image obtained after decomposition by the Retinex network. The multi-scale enhancement network is responsible for further improving the quality of the image processed by the reflection image enhancement network. The goal of this module is to generate an optimized uneven low-light image enhancement model; Finally, there is the image output module, which is responsible for displaying the enhanced high-quality image. This module uses electronic devices to output the processed high-quality image so that users can observe and evaluate the final result. These four modules work together to form a process, which ultimately achieves the enhancement of uneven low-light images through preprocessing, training and output. Applied to a non-uniform low-illumination image enhancement method as described in any one of claims 1 to 7.

Images