A Low-Light Image Enhancement Method Based on DenseNet and GAN

Abstract

The present invention discloses a low-light image enhancement method based on DenseNet and GAN, belonging to the field of image information processing. The specific steps of the present invention are as follows: Step 1: Construct a generator network using the DenseNet framework; Step 2: Construct a discriminator network using the PatchGAN framework; Step 3: Optimize the network weights based on the game between the generator network and the discriminator network, and continuously adjust the enhancement effect of the low-light image to obtain the final image. Aiming at the practical problems such as low signal-to-noise ratio, unsatisfactory resolution and illumination level in the collected images under low-light environments, this method uses DenseNet to improve the classical GAN model, contributing to a reasonable solution to the problem of low-light image enhancement and having high academic value.

Description

Technical Field

[0001] The present invention belongs to the field of image information processing, and particularly relates to a low-light image enhancement method based on DenseNet and GAN. Background Art

[0002] During the image acquisition process, the lighting conditions of the scene often become an important factor affecting the image quality. In modern social production and life, it has become increasingly convenient and fast for people to acquire images. However, due to insufficient lighting conditions, the recognition rate of low-light images is not high, which greatly affects the use value and causes great difficulties for subsequent image processing, target recognition, semantic segmentation and other tasks.

[0003] Currently, deep learning is developing rapidly and has extensive applications and good effects in computer vision. Deep learning models can use a large amount of data to complete the training of the deep model. For example, by inputting a large number of images with different illuminations and using an autoencoder or a convolutional neural network for training, the effect of enhancing low-light images can be achieved.

[0004] In order to improve the processing effect of low-light image enhancement and obtain high-quality images, the present invention utilizes the training characteristics and learning ability of the generative adversarial network, and proposes a low-light image enhancement method based on a dense connection network and a generative adversarial network. This method uses the dense connection network (DenseNet) as a framework to build a generator network, and its advantage lies in having very good anti-overfitting performance, especially suitable for tasks with relatively scarce training data. In addition, the present invention uses the Markov discriminator (PatchGAN) as the discriminator network to supervise the generator network, thereby improving the ability of the generator network to generate high-quality images.

[0005] The following is a partial Chinese-English comparison:

[0006] Dense Connection Network (DenseNet)

[0007] Generative Adversarial Network (GAN)

[0008] Batch Normalization Technique (BatchNorm)

[0009] Rectified Linear Unit (ReLu)

[0010] Hyperbolic Tangent Function (Tanh)

[0011] Concatenation

[0012] Mean Squared Error (MSE). Summary of the Invention

[0013] Objective of the invention: To provide a low-light image enhancement method based on DenseNet and GAN with better effects. For specific objectives, see multiple substantial technical effects in the specific implementation part.

[0014] To achieve the above objective, the present invention adopts the following technical solutions:

[0015] Technical solution: A low-light image enhancement method based on a dense connection network and a generative adversarial network, comprising the following steps:

[0016] Step 1, constructing a generator network using the DenseNet framework;

[0017] (1) Constructing 4 upsampling layers with 4×4 convolutional kernels to extract features from the input low-light image, where the number of convolutional kernels in each layer is (1024, 512, 256, 128). After convolution in each layer, BatchNorm is used to standardize the data, and then ReLu is used as the activation function to perform upsampling on the input data, and finally it is passed into DenseNet. The mathematical expression for upsampling feature extraction is:

[0018] G Li (X) = C Li *X (1)

[0019] Among them, X is the input low-light image, G Li represents the i-th group of feature maps output by the L-th layer, "*" represents the convolution operation, C Li represents the i-th group of convolutional kernels of the L-th layer, i = 1, 2, 3, 4.

[0020] (2) When designing the DenseNet structure, on the one hand, the characteristics of the small sample space are considered, and on the other hand, to avoid too many parameters and overfitting caused by too many layers, the present invention designs DenseNet into four layers. Among them, Layer1-Layer3 all use 5×5 convolutional kernels and the number of channels is 128. After convolution between layers, BatchNorm is used for data standardization, and ReLu activation function is used. Layer4 uses 5×5 convolutional kernels and the number of channels is 3. After convolution, Tanh is used as the activation function. The entire DenseNet framework uses 3 concatenations to adjust feature maps of different sizes to a unified number of channels for connection.

[0021] The mapping function expression of DenseNet is:

[0022] X L = H L ([X0, X1,..., X L-1 ) (2)

[0023] Among them, L represents the total number of layers, and "[]" means combining all the output feature maps from the X0th layer to the X L-1 layer according to the number of channels, and H L represents performing non-linear processing using BatchNorm + ReLu + Conv. X on the left side of the equal sign L represents the result after adopting H L processing.

[0024] Step 2: Construct a discriminator network using the PatchGAN framework;

[0025] In the present invention, when constructing the discriminator network, a PatchGAN discriminator is used to replace the traditional binary classification discriminator. On the one hand, this discriminator can output an N×N matrix, and judge the real image and the image generated by the generator through the calculation of the matrix mean. This discriminator fully considers the features of different regions of the image, pays attention to the capture and consideration of image detail information, and helps to improve the image quality; on the other hand, this discriminator calculates for small-sized image patches, which improves the network convergence speed. The present invention designs a convolutional neural network with four fully convolutional layers. Except for the last convolutional layer, BatchNorm is used to standardize the data after convolution between each of the other three layers, and LeakyReLu is used as the activation function.

[0026] Step 3: Execute Step 3.1 to construct the loss function, execute Step 3.2 to perform the game optimization of the network weights between the generator network and the discriminator network, and continuously adjust the enhancement effect of the low-illumination image to obtain the final image;

[0027] Step 3.1: Construct the loss function:

[0028] In the training process of the generator network and the discriminator network of the present invention, the mean square error (MSE) index is used as the loss function. Among them, for the generator network, MSE is used to calculate the error value between the image after generating illumination enhancement and the normal illumination image, while for the discriminator network, MSE is used to calculate the error value between the image processed by the discriminator network and the normal illumination image. The mathematical expressions of the loss functions of the generator network and the discriminator network are respectively:

[0029]

[0030]

[0031] Among them, D and G respectively represent the discriminator network and the generator network, and L G and L Drespectively represent the error values of the generator network and the discriminator network, N is the number of training samples, and Y i represents the value of the i-th group of real normal illumination images, T i represents the i-th group of outputs generated by the generator network, B i represents the i-th group of outputs generated by the discriminator network.

[0032] Step 3.2: Perform a game between the generator network and the discriminator network to optimize the network weights, and continuously adjust the enhancement effect of the low-illumination image to obtain the final image;

[0033] Construct an evaluation function:

[0034]

[0035] The larger the evaluation function value, the better the corresponding image enhancement effect. Therefore, it is necessary to perform alternating iterations on the generator network D and the discriminator network G. That is, within a certain period of time, the parameters in the discriminator network G remain unchanged to optimize the generator network D, and within another period of time, the parameters in the generator network G remain unchanged to optimize the discriminator network D. The iterative formulas for the generator network G and the generator network D are respectively:

[0036]

[0037]

[0038] As the number of iterations increases, the generated image with enhanced illumination basically coincides with the normal illumination image, making the discriminator network in a Nash equilibrium, that is, the probability of accurate discrimination is 0.5 and accurate judgment cannot be completed. The iteration ends, and the generated image with enhanced illumination is the final image.

[0039] Adopting the above technical solutions, the present invention has the following beneficial effects: For the current mainstream low-illumination image enhancement methods, the image with enhanced illumination obtained by the method of the present invention is not only very close to the normal illumination image in terms of brightness, but also well realizes the retention of image detail information and color restoration. The enhanced image is more natural and has a better intuitive visual effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] To further illustrate the present invention, the following is further described in conjunction with the drawings:

[0041] Figure 1 is the flow chart of the method of the present invention;

[0042] Figures 2 - 3 is the source image used in the simulation experiment; Figure 2 is the low-illumination image; Figure 3 is the corresponding normal illumination image.

[0043] Figures 4 - 11 are the effect simulation diagrams of the simulation experiment; among them, Figure 4 are the simulation results of the HE method; Figure 5 are the simulation results of the MSRCR method; Figure 6 are the simulation results of the Gou method; Figure 7 are the simulation results of the Ying method; Figure 8 are the simulation results of the Lore method; Figure 9 are the simulation results of the Ma method; Figure 10 are the simulation results of the Li method; Figure 11 are the simulation results of the method of the present invention. Specific embodiments

[0044] The present invention will be further clarified below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0045] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0046] This patent provides multiple parallel solutions. For different expressions, they belong to improved solutions or parallel solutions based on the basic solution. Each solution has its own unique features. In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0047] A low-light image enhancement method based on a dense connection network and a generative adversarial network of the present invention specifically includes the following steps:

[0048] Step 1, construct a generator network using the DenseNet framework;

[0049] (1) Four upsampling layers with 4×4 convolutional kernels are constructed to extract features from the input low - illumination image. The number of convolutional kernels in each layer is (1024, 512, 256, 128) respectively. After convolution in each layer, BatchNorm is used to standardize the data, and then ReLu is used as the activation function to perform upsampling on the input data, and finally it is passed into DenseNet. The mathematical expression for upsampling feature extraction is:

[0050] G Li (X)=C Li *X (1)

[0051] Among them, X is the input low - illumination image, G Li represents the i - th group of feature maps output by the L - th layer, "*" represents the convolution operation, C Li represents the i - th group of convolutional kernels of the L - th layer, and i = 1, 2, 3, 4.

[0052] (2) When designing the DenseNet structure, on the one hand, the characteristics of the small - sample space are considered, and on the other hand, in order to avoid too many parameters and overfitting caused by too many layers, the present invention designs DenseNet into four layers. Among them, the convolutional kernels of Layer1 - Layer3 are all 5×5 and the number of channels is 128. After convolution between layers, BatchNorm is used for data standardization, and the ReLu activation function is used. Layer4 uses a 5×5 convolutional kernel and the number of channels is 3. After convolution, Tanh is used as the activation function. The entire DenseNet framework uses 3 concatenations to adjust feature maps of different sizes to the same number of channels for connection.

[0053] The mapping function expression of DenseNet is:

[0054] X L =H L ([X0,X1,...,X L-1 ) (2)

[0055] Among them, L represents the total number of layers, "[]" represents combining all the output feature maps from the X0 - th layer to the X L-1 - th layer according to the number of channels, H L represents performing non - linear processing using BatchNorm + ReLu + Conv, and X L on the left side of the equal sign represents the result after processing using H L .

[0056] Step 2: Construct a discriminator network using the PatchGAN framework;

[0057] In constructing the discriminator network, the present invention uses a PatchGAN discriminator to replace the traditional binary classifier. On the one hand, this discriminator can output an N×N matrix and judge the real image and the image generated by the generator through matrix mean calculation. This discriminator fully considers the features of different regions of the image, pays attention to the capture and consideration of image detail information, and helps to improve the image quality. On the other hand, this discriminator calculates for small-size image patches, which improves the network convergence speed. The present invention designs a convolutional neural network with four fully convolutional layers. Except for the last convolutional layer, BatchNorm is used to standardize the data after convolution between each of the other three layers, and LeakyReLu is used as the activation function.

[0058] Step 3: Execute Step 3.1 to construct the loss function, and execute Step 3.2 to perform the game between the generator network and the discriminator network to optimize the network weights, and continuously adjust the enhancement effect of the low-illumination image to obtain the final image.

[0059] Step 3.1: Construct the loss function:

[0060] In the training process of the generator network and the discriminator network of the present invention, the mean square error (MSE) index is used as the loss function. Among them, for the generator network, MSE is used to calculate the error value between the image with enhanced illumination and the normal-illumination image, while for the discriminator network, MSE is used to calculate the error value between the image processed by the discriminator network and the normal-illumination image. The mathematical expressions of the loss functions of the generator network and the discriminator network are respectively:

[0061]

[0062]

[0063] Among them, D and G respectively represent the discriminator network and the generator network, L G and L D respectively represent the error value of the generator network and the error value of the discriminator network, N is the number of training samples, Y i represents the value of the i-th group of real normal-illumination image, T i represents the i-th group of output generated by the generator network, and B i represents the i-th group of output generated by the discriminator network.

[0064] Step 3.2: Perform the game between the generator network and the discriminator network to optimize the network weights, and continuously adjust the enhancement effect of the low-illumination image to obtain the final image.

[0065] Construct the evaluation function:

[0066]

[0067] The larger the evaluation function value is, the better the corresponding image enhancement effect is. Therefore, it is necessary to perform alternating iterations on the generator network D and the discriminator network G. That is, within a certain period of time, the parameters in the discriminator network G remain unchanged to optimize the generator network D, and within another period of time, the parameters in the generator network G remain unchanged to optimize the discriminator network D. The iterative formulas for the generator network G and the discriminator network D are as follows:

[0068]

[0069]

[0070] As the number of iterations continues to increase, the generated image with enhanced illumination basically coincides with the normal illumination image, making the discriminator network in a Nash equilibrium, that is, the probability of accurate discrimination is 0.5 and accurate judgment cannot be completed, and the iteration ends. Then the generated image with enhanced illumination is the final image.

[0071] Embodiment:

[0072] The simulation experiment platform for the method of the present invention is a personal PC with a configuration of Intel(R) Core(TM) i5-4250U CPU 1.90GHz and 4GB of memory. The simulation software is Matlab 2014b. The technical solution of the present invention is as Figure 1 shown. In order to better understand the technical solution of the present invention, two source images (low-illumination image + normal-illumination image) are selected in this embodiment for simulation experiments. Following the technical solution of the present invention, a low-illumination image enhancement method based on a dense connection network and a generative adversarial network in this embodiment specifically includes the following steps:

[0073] Step 1, construct a generator network using the DenseNet framework;

[0074] (1) Construct 4 upsampling layers with 4×4 convolutional kernels to perform feature extraction on the input low-illumination image, where the number of convolutional kernels in each layer is (1024, 512, 256, 128). After convolution in each layer, BatchNorm is used to standardize the data, and then ReLu is used as the activation function to perform upsampling processing on the input data, and finally it is passed into DenseNet. The mathematical expression for upsampling feature extraction is:

[0075] G Li (X) = C Li *X (1)

[0076] where X is the input low-illumination image, G LiRepresents the i-th group of feature maps of the output of the L-th layer. "*" represents the convolution operation, and C Li Represents the i-th group of convolution kernels of the L-th layer, where i = 1, 2, 3, 4.

[0077] (2) When designing the DenseNet structure, on the one hand, the characteristics of the small sample space are considered, and on the other hand, in order to avoid too many parameters and overfitting caused by too many layers, the present invention designs DenseNet into four layers. Among them, the convolution kernels of Layer 1 - Layer 3 are all 5×5 and the number of channels is 128. After convolution between layers, BatchNorm is used for data normalization processing, and the ReLu activation function is used. Layer 4 uses a 5×5 convolution kernel and the number of channels is 3. After convolution, Tanh is used as the activation function. The entire DenseNet framework uses 3 concatenations to adjust the feature maps of different sizes to a unified number of channels for connection.

[0078] The mapping function expression of DenseNet is:

[0079] X L = H L ([X0, X1,..., X L-1 ) (2)

[0080] Among them, L represents the total number of layers. "[]" represents combining all the output feature maps of the X0-th layer to the X L-1 -th layer according to the number of channels. H L represents using BatchNorm + ReLu + Conv for non-linear processing. X on the left side of the equal sign L represents the result after being processed by H L .

[0081] Step 2, construct a discriminator network using the PatchGAN framework;

[0082] When constructing the discriminator network, the present invention uses a PatchGAN discriminator to replace the traditional binary classification discriminator. On the one hand, this discriminator can output an N×N matrix and judge the real image and the image generated by the generator through matrix mean calculation. This discriminator fully considers the characteristics of different regions of the image, pays attention to the capture and consideration of image detail information, and helps to improve the image quality. On the other hand, this discriminator calculates for small-sized image patches, which improves the network convergence speed. The present invention designs a convolutional neural network with four fully convolutional layers. Except for the last convolutional layer, BatchNorm is used to normalize the data after convolution between the other three layers, and LeakyReLu is used as the activation function.

[0083] Step 3: Execute Step 3.1 to construct a loss function, and execute Step 3.2 to perform game optimization on the network weights between the generator network and the discriminator network, continuously adjust the enhancement effect of the low-illumination image, and thus obtain the final image;

[0084] Step 3.1: Construct a loss function:

[0085] During the training process of the generator network and the discriminator network of the present invention, the mean square error (MSE) index is used as the loss function. Among them, for the generator network, MSE is used to calculate the error value between the image with enhanced illumination and the normal-illumination image, and for the discriminator network, MSE is used to calculate the error value between the image processed by the discriminator network and the normal-illumination image. The mathematical expressions of the loss functions of the generator network and the discriminator network are respectively:

[0086]

[0087]

[0088] Among them, D and G respectively represent the discriminator network and the generator network, L G and L D respectively represent the error value of the generator network and the error value of the discriminator network, N is the number of training samples, Y i represents the i-th group of true normal-illumination image values, T i represents the i-th group of outputs generated by the generator network, and B i represents the i-th group of outputs generated by the discriminator network.

[0089] Step 3.2: Perform game optimization on the network weights between the generator network and the discriminator network, continuously adjust the enhancement effect of the low-illumination image, and thus obtain the final image;

[0090] Construct an evaluation function:

[0091]

[0092] The larger the evaluation function value, the better the corresponding image enhancement effect. Therefore, it is necessary to perform alternating iterations on the generator network D and the discriminator network G, that is, within a certain period of time, the parameters in the discriminator network G remain unchanged to optimize the generator network D, and within another period of time, the parameters in the generator network G remain unchanged to optimize the discriminator network D. The iteration formulas of the generator network G and the generator network D are respectively:

[0093]

[0094]

[0095] As the number of iterations continues to increase, the generated image with enhanced illuminance basically coincides with the normal-illuminance image, causing the discriminator network to be in a Nash equilibrium, that is, the probability of accurate discrimination is 0.5 and accurate judgment cannot be completed. The iteration ends, and the generated image with enhanced illuminance is the final image.

[0096] Simulation comparison experiment:

[0097] To verify the effectiveness of the method of the present invention, the following is a set of simulation experiments to verify that compared with various current mainstream low-illuminance image enhancement methods, the method of the present invention has better experimental results:

[0098] Following the technical solution of the present invention, for a set of source images to be processed, this set of source images includes a low-illuminance image (see Figure 2 ) and the corresponding normal-illuminance image (see Figure 3 ). At the same time, several representative methods are selected, including the HE method, the MSRCR method, the Gou method, the Ying method, the Lore method, the Ma method, the Li method, and the corresponding method of the present invention for comparison.

[0099] Figures 4 - 11 The simulation experiment results of the eight methods are given, which show that the image with enhanced illuminance obtained by the method of the present invention is not only very close to the brightness of the normal-illuminance image in terms of brightness, but also well realizes the retention of image detail information and color restoration. The enhanced image is more natural and has a better intuitive visual effect. In addition, the present invention also selects the Peak Signal-to-Noise Ratio (PSNR), the Structural Similarity Index (SSIM), and the Average Running Time (ART) as the objective quality evaluation indicators for the eight methods. Table 1 gives the objective evaluation results of the corresponding result images of the eight low-illuminance image enhancement methods in the simulation experiment. It should be noted that the bold font in the table represents the optimal result under the same objective index. Obviously, compared with the other seven methods, the method of the present invention performs best in both the PSNR and SSIM indexes. This result shows that the finally enhanced illuminance image generated by the method of the present invention has a more ideal intuitive visual effect. Table 2 gives the objective evaluation results of the average running time indexes corresponding to the eight low-illuminance image enhancement methods. Among the eight methods, the Gou method and the Li method have the shortest average running time, and the method of the present invention ranks third. It should be noted that although the average running time of the method of the present invention is slightly higher than the first two methods, the finally enhanced illuminance image obtained by the method of the present invention has a better intuitive visual effect and objective evaluation results than the first two methods.

[0100] Table 1 Objective Evaluation Results of PSNR and SSIM Metrics for Eight Image Enhancement Methods

[0101]

[0102]

[0103] Table 2 Objective Evaluation Results of Average Running Time Metrics for Eight Image Enhancement Methods (Unit: seconds)

[0104]

[0105] Generally speaking, considering the intuitive visual effect, objective evaluation index values, and average running time comprehensively, the method of the present invention has significant advantages compared with various current mainstream low-light image enhancement methods.

[0106] The present invention discloses a low-light image enhancement method based on a DenseNet (Dense Convolutional Network) and a Generative Adversarial Network (GAN), belonging to the field of image information processing. The specific steps of the present invention are as follows: Step 1: Construct a generator network using the DenseNet framework; Step 2: Construct a discriminator network using the PatchGAN framework; Step 3: Optimize the network weights based on the game between the generator network and the discriminator network, and continuously adjust the enhancement effect of the low-light image to obtain the final image. Aiming at the practical problems that the captured images in low-light environments usually have low signal-to-noise ratio, unsatisfactory resolution and illumination level, etc., the DenseNet is used to improve the classical GAN model in this method, which contributes to the reasonable solution of the low-light image enhancement problem and has high academic value.

[0107] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the claimed invention.

Claims

1. A low-light image enhancement method based on DenseNet and GAN, characterized in that It includes the following steps: Step 1: Construct a generator network using the DenseNet framework; Step 2: Construct a discriminator network using the PatchGAN framework; Step 3: Execute Step 3.1 to construct a loss function, and execute Step 3.2 to perform game optimization on the network weights between the generator network and the discriminator network, continuously adjusting the enhancement effect of the low-illumination image to obtain the final image; The implementation method of Step 1: (1) Construct 4 upsampling layers with 4×4 convolutional kernels to extract features from the input low-illumination image. The number of convolutional kernels in each layer is (1024, 512, 256, 128). After each layer of convolution, batch normalization technology is used to standardize the data, and then the rectified linear unit function is used as the activation function to perform upsampling on the input data, and finally it is passed into DenseNet. The mathematical expression of upsampling feature extraction is: G Li (X) = C Li *X(1) Among them, X is the input low-light image, and G Li represents the i-th group of feature maps output by the L-th layer, "*" represents the convolution operation, and C Li represents the i-th group of convolution kernels of the L-th layer, where i = 1, 2, 3, 4; (2) When designing the DenseNet structure, DenseNet is designed as four layers. Among them, Layer 1 - Layer 3 all use 5×5 convolutional kernels and the number of channels is 128. After convolution between layers, BatchNorm is used for data normalization, and the ReLu activation function is used. Layer 4 uses a 5×5 convolutional kernel and the number of channels is 3. After convolution, the hyperbolic tangent function is used as the activation function. The entire DenseNet framework uses 3 concatenations to adjust feature maps of different sizes to the same number of channels for connection; The mapping function expression of DenseNet is: X L = H L ([X0,X1,...,X L-1 ) (2) Among them, L represents the total number of layers, and "[]" means combining all the output feature maps from the X0th layer to the X L-1 th layer according to the number of channels. H L represents the use of BatchNorm + ReLu + Conv for non-linear processing. X on the left side of the equal sign L represents the result after adopting H L processing; The implementation method of Step 2: In the present invention, a PatchGAN discriminator is used to replace the traditional binary classification discriminator when constructing the discriminator network. On the one hand, this discriminator can output an N×N matrix, and judge the real image and the image generated by the generator through matrix mean calculation. On the other hand, this discriminator calculates for small-size image patches, improving the network convergence speed. The present invention designs a convolutional neural network with four fully convolutional layers. Except for the last convolutional layer, BatchNorm is used to standardize the data after convolution between the other three layers, and LeakyReLu is used as the activation function.

2. The low - illumination image enhancement method based on DenseNet and GAN according to claim 1, wherein The specific steps of Step 3.1 in Step 3 are as follows: In the training process, both the generator network and the discriminator network in the steps use the mean square error index as the loss function. Among them, for the generator network, the mean square error is used to calculate the error value between the image after illumination enhancement and the normal illumination image. For the discriminator network, MSE is used to calculate the error value between the image processed by the discriminator network and the normal illumination image. The mathematical expressions of the loss functions of the generator network and the discriminator network are respectively: Among them, D and G respectively represent the discriminator network and the generator network, L G and L D respectively represent the error value of the generator network and the error value of the discriminator network. N is the number of training samples, and Y i represents the value of the i-th group of real normal illuminance images, and T i represents the i-th group of outputs generated by the generator network, and B i represents the i-th group of outputs generated by the discriminator network.

3. The low - illumination image enhancement method based on DenseNet and GAN according to claim 1, characterized in that, The specific steps of Step 3.2 in Step 3 are as follows: Construct an evaluation function: The larger the evaluation function value is, the better the corresponding image enhancement effect is. Therefore, it is necessary to perform alternating iterations on the generator network D and the discriminator network G. That is, within a certain period of time, the parameters in the discriminator network G remain unchanged to optimize the generator network D, and within another period of time, the parameters in the generator network G remain unchanged to optimize the discriminator network D. The iterative formulas for the generator network G and the discriminator network D are as follows: As the number of iterations continues to increase, the generated image with enhanced illumination basically coincides with the normal illumination image, making the discriminator network in a Nash equilibrium, that is, the probability of accurate discrimination is 0.5 and accurate judgment cannot be completed. When the iteration ends, the generated image with enhanced illumination is the final image.

Images