Low illumination image enhancement method based on convolutional neural network

Abstract

The present invention discloses a low-light image enhancement method based on a convolutional neural network, which relates to the field of image enhancement technology and includes the following steps: S1, obtaining a normal-light image, using an image reconstruction model to artificially synthesize a low-light image, and using the synthesized low-light image as an image set; S2, dividing the obtained image set into a first training set and a first test set according to corresponding proportions, and training an image detection model through the first training set to obtain a trained image detection model, and performing a performance test on the trained image detection model through the first test set, wherein the input end of the image detection model is an image, and the output end of the image detection model is an image detected as low-light. By using the image reconstruction model to artificially synthesize low-light images, the present invention can better control lighting conditions and other influencing factors, so as to conduct research and processing on specific low-light situations, and can better simulate low-light images in different scenarios.

Description

Technical Field

[0001] The present invention relates to the field of image enhancement technology, and in particular to a low-illumination image enhancement method based on a convolutional neural network. Background Art

[0002] Low-light images refer to images captured or acquired in dark or insufficient light conditions, resulting in low brightness. In this case, the details and colors in the image are limited, affecting the image quality and clarity. Low-light images exhibit characteristics such as weak contrast, blurred details, and high noise. When capturing images in low-light environments, image enhancement technology is often required to improve image quality and make them clearer and brighter.

[0003] Low-light images captured in actual environments are often limited by various factors such as lighting conditions and camera performance. It is difficult to ensure that the images have a certain quality and diversity. In addition, the annotation of low-light images usually requires professionals, and such annotation data is difficult to collect and obtain. In addition, existing image enhancement methods cannot automatically detect low-light images. Manual detection of low-light images will become very time-consuming and cumbersome, and the accuracy may be limited, making it impossible to quickly identify and filter out low-light images. This may lead to a decrease in work efficiency when processing large amounts of image data, which may result in poor training effects of image enhancement models. Summary of the Invention

[0004] The purpose of the present invention is to provide a low-light image enhancement method based on convolutional neural network to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a low-light image enhancement method based on a convolutional neural network, comprising the following steps:

[0006] S1. Obtain a normal illumination image, use an image reconstruction model to artificially synthesize a low illumination image, and use the synthesized low illumination image as an image set;

[0007] S2. Divide the acquired image set into a first training set and a first test set according to corresponding proportions, and train the image detection model using the first training set to obtain a trained image detection model, and perform a performance test on the trained image detection model using the first test set, where the input end of the image detection model is the image, and the output end of the image detection model is the image detected as low-light;

[0008] S3. Dividing the images detected as low-illuminance into a second training set and a second test set, and training the image enhancement model using the second training set to obtain a trained image enhancement model, and performing a performance test on the trained image enhancement model using the second test set, where the input end of the image enhancement model is the low-illuminance image, and the output end of the image enhancement model is the enhanced image;

[0009] S4. The image enhancement model includes an image decomposition model, a reflectance map refinement model, and an illumination map enhancement model. The image decomposition model is trained using the second training set to obtain a trained image decomposition model. The input of the image decomposition model is the low-light image, and the output of the image decomposition model is the reflectance map and the illumination map.

[0010] S5. Training a reflectivity map refinement model using the reflectivity map to obtain a trained reflectivity map refinement model, wherein an input end of the reflectivity map refinement model is the reflectivity map, and an output end of the reflectivity map refinement model is the refined reflectivity map;

[0011] S6. Training a light map enhancement model using the light map to obtain a trained light map enhancement model, wherein an input end of the light map enhancement model is the light map, and an output end of the light map enhancement model is the enhanced light map;

[0012] S7. Perform gradient weighting on the refined reflectance map and the enhanced illumination map to obtain an enhanced image.

[0013] Preferably, the image detection model adopts a convolutional neural network, which includes a convolution layer, a pooling layer and a fully connected layer. The convolution layer, the pooling layer and the fully connected layer all have 3 layers. The first layer of the image detection model is a convolution layer, the number of neurons is 64, the convolution kernel size is 7*7, and the step size is 1. The second layer is a pooling layer, the convolution kernel size is 2*2, and the step size is 2. The third layer is a convolution layer, the number of neurons is 128, the convolution kernel size is 5*5, and the step size is 1. The fourth layer is a pooling layer, the convolution kernel size is 2*2, and the step size is 2. The fifth layer is a convolutional layer with 256 neurons, a convolution kernel size of 3*3, and a step size of 1. The sixth layer is a pooling layer with a convolution kernel size of 2*2 and a step size of 2. The seventh and eighth layers are fully connected layers with 512 neurons each. The ninth layer is a fully connected layer with 2 neurons. The first, third, fifth, seventh, and eighth layers are all equipped with activation functions, and the activation functions are all relu functions. The first 6 layers of the image detection model are used to extract image features, and the last three layers of the image detection model are used for image classification.

[0014] Preferably, the loss function of the image detection model is:

[0015]

[0016] Among them, K represents the number of images, M represents the number of image categories, and w j Expressed as the class weight of the jth class, p ij represents the true probability that the i-th image belongs to the j-th category, p′ ij represents the predicted probability that the i-th image belongs to the j-th class, and N represents the total number of samples in the data set, j represents the number of samples in the jth class, w j It is expressed as the class weight of the jth class.

[0017] Preferably, the image decomposition model adopts a convolutional neural network, and the image decomposition model includes a convolution layer and a merging layer. The first layer of the image decomposition model includes three convolution layers, the number of neurons in the three convolution layers are 12, 20 and 32 respectively, and the convolution kernel sizes of the three convolution layers are 3*3, 5*5 and 7*7 respectively. Different image features are extracted through the first layer, and the second layer is a merging layer. Each feature map is split according to the channel dimension through the second layer, and processed separately, and the processed feature maps are spliced ​​on the channel dimension to form The new feature map, the third and fourth layers are both convolution layers, the number of neurons is 64, the convolution kernel size is 3*3, the image features are refined through the third and fourth layers, the fifth layer is a convolution layer, the number of neurons is 4, the convolution kernel size is 3*3, the image features are decomposed into reflectance map and illumination map through the fifth layer, and the reflectance map and illumination map are output. Finally, the fifth layer sets the Sigmoid activation function, and the pixel values ​​of the output reflectance map and illumination map are normalized to [0,1] through the Sigmoid activation function.

[0018] Preferably, the loss function of the image decomposition model is:

[0019] L=L s +λ r L r +λ l L l

[0020]

[0021] L r =||R low -R normal ||2

[0022]

[0023] Among them, L s Represents the loss function of the reconstructed image after image restoration, L r Represents the loss function for the reflectivity map to remain unchanged, Ll represents the loss function of the smooth illumination map, λ r and λ l The constants representing the constancy of the balanced reflectance map and the smoothness of the illumination map, R i represents the normal lighting image vector of the i-th category, S j represents the j-th normal lighting image vector, L i represents the light map vector, e -λ▽R represents smoothness constraint, R normal Represents the reflectivity map, R low Reflectance map representing a low-light image.

[0024] Preferably, the reflectance map refinement model adopts a convolutional neural network, and the reflectance map refinement model includes a convolution layer. The first layer of the reflectance map refinement model is a convolution layer, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is a PReLu function. The coarsely extracted reflectance map is convolved through the first layer, the second layer is a convolution layer, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is a PReLu function. The reflectance map after convolution of the first layer is reconstructed through the second layer, the third layer is a convolution layer, the number of neurons is 3, the convolution kernel is 3*3, and the activation function is a Sigmoid function. The reflectance map reconstructed in the second layer is restored in detail through the third layer, and the pixel value of the output reflectance map is normalized to [0,1] through the Sigmoid activation function.

[0025] Preferably, the loss function of the reflectance map refinement model is:

[0026] L=||P′-P||2

[0027] P=R normal -R low

[0028] Where P′ represents R normal and R low The residual, P represents the true residual, R normal Represents the reflectivity map, R low Reflectance map representing a low-light image.

[0029] Preferably, the illumination map enhancement model adopts a convolutional neural network, and the illumination map enhancement model includes a convolution layer and a connection layer. The first three layers of the reflectance map refinement model are all convolution layers, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is PReLu function. The first three layers are encoded and downsampled to extract the global illumination map features. The middle three layers are all convolution layers, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is PReLu function. The enhanced features of the illumination map are obtained through the middle three layers. The seventh layer is a connection layer, and jump connection is performed through the seventh layer to combine multi-layer features. The eighth layer is a convolution layer, the number of neurons is 1, the convolution kernel is 3*3, and the activation function is Sigmoid function. The pixel value of the output enhanced illumination map is normalized to [0,1] through the Sigmoid activation function.

[0030] Preferably, the loss function of the illumination map enhancement model is:

[0031] L=L1+λ1L2+λ2L3

[0032]

[0033] L3=||L′ low -L normal ||2

[0034] Among them, L represents the total loss function, L1 represents the reconstruction loss function, L2 represents the loss function based on the characteristics of the illumination map, L3 represents the consistency loss function between the enhanced illumination map and the normal illumination image, and L′ low represents the enhanced illumination map, S normal represents the normal lighting image, λ r and λ l Constants representing the constancy of light map properties and the consistency of the light map, respectively.

[0035] Preferably, in S7, performing gradient weighting on the refined reflectance map and the enhanced illumination map to obtain an enhanced image comprises the following steps:

[0036] S701, taking the refined reflectivity map and the enhanced illumination map as input;

[0037] S702, performing gradient calculation on the reflectivity map and the illumination map, obtaining the gradient magnitude and direction of the reflectivity map and the illumination map, and multiplying the gradient magnitude by a weight coefficient;

[0038] S703, weighting the two gradient images according to the magnitude and direction of the gradient amplitude;

[0039] S704: Fusing the weighted gradient image with the original low-light image using a weighted superposition method to obtain an enhanced image.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. By using an image reconstruction model to artificially synthesize low-light images, the present invention can better control lighting conditions and other influencing factors to study and process specific low-light situations, thereby better simulating low-light images in different scenarios, thereby better evaluating and improving low-light image enhancement methods based on convolutional neural networks. In addition, the image detection model can automatically detect low-light conditions on images without manual intervention, which greatly improves efficiency and accuracy and saves human resources. In addition, the image detection model can perform rapid detection in real-time scenarios and can promptly detect and respond to low-light images, which is very important for some time-sensitive applications, such as video surveillance and autonomous driving.

[0042] 2. The present invention can effectively improve the quality of low-light images through image decomposition, reflection map refinement and illumination map enhancement processes through the image enhancement model, making low-light images clearer and brighter, reducing noise and artifacts, and increasing details and contrast. Low-light images often lose some details and color information. The image enhancement model can restore and enhance this lost information through processing steps such as reflection map refinement and illumination map enhancement, making the image more realistic and visual. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic flow chart of the overall method provided by an embodiment of the present invention;

[0044] Figure 2 A flowchart of the method in S7 provided in an embodiment of the present invention;

[0045] Figure 3 A flowchart of the overall method provided by an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of a low-light image provided by an embodiment of the present invention;

[0047] Figure 5 A schematic diagram of an enhanced image provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0049] See also Figure 1-5 The present invention provides a technical solution: a low-light image enhancement method based on a convolutional neural network, comprising the following steps:

[0050] S1. Obtain a normal illumination image, use an image reconstruction model to artificially synthesize a low illumination image, and use the synthesized low illumination image as an image set;

[0051] S2. Divide the acquired image set into a first training set and a first test set according to corresponding proportions, and train the image detection model using the first training set to obtain a trained image detection model, and perform a performance test on the trained image detection model using the first test set, where the input end of the image detection model is the image, and the output end of the image detection model is the image detected as low-light;

[0052] S3. Dividing the images detected as low-illuminance into a second training set and a second test set, and training the image enhancement model using the second training set to obtain a trained image enhancement model, and performing a performance test on the trained image enhancement model using the second test set, where the input end of the image enhancement model is the low-illuminance image, and the output end of the image enhancement model is the enhanced image;

[0053] S4. The image enhancement model includes an image decomposition model, a reflectance map refinement model, and an illumination map enhancement model. The image decomposition model is trained using the second training set to obtain a trained image decomposition model. The input of the image decomposition model is the low-light image, and the output of the image decomposition model is the reflectance map and the illumination map.

[0054] S5. Training a reflectivity map refinement model using the reflectivity map to obtain a trained reflectivity map refinement model, wherein an input end of the reflectivity map refinement model is the reflectivity map, and an output end of the reflectivity map refinement model is the refined reflectivity map;

[0055] S6. Training a light map enhancement model using the light map to obtain a trained light map enhancement model, wherein an input end of the light map enhancement model is the light map, and an output end of the light map enhancement model is the enhanced light map;

[0056] S7. Perform gradient weighting on the refined reflectance map and the enhanced illumination map to obtain an enhanced image.

[0057] The image detection model uses a convolutional neural network. The image detection model includes convolutional layers, pooling layers, and fully connected layers. The convolutional layers, pooling layers, and fully connected layers all have 3 layers. The first layer of the image detection model is a convolutional layer with 64 neurons, a convolution kernel size of 7*7, and a step size of 1. The second layer is a pooling layer with a convolution kernel size of 2*2 and a step size of 2. The third layer is a convolutional layer with 128 neurons, a convolution kernel size of 5*5, and a step size of 1. The fourth layer is a pooling layer with a convolution kernel size of 2*2 and a step size of 2. The fifth layer is a convolutional layer with 128 neurons, a convolution kernel size of 5*5, and a step size of 1. The first layer is a convolutional layer with 256 neurons, a convolution kernel size of 3*3, and a stride of 1. The sixth layer is a pooling layer with a convolution kernel size of 2*2 and a stride of 2. The seventh and eighth layers are fully connected layers with 512 neurons each. The ninth layer is a fully connected layer with 2 neurons. The first, third, fifth, seventh, and eighth layers are all equipped with activation functions, and the activation functions are all relu functions. The first six layers of the image detection model are used to extract image features, and the last three layers of the image detection model are used for image classification.

[0058] The loss function of the image detection model is:

[0059]

[0060] Among them, K represents the number of images, M represents the number of image categories, and w j Expressed as the class weight of the jth class, p ij represents the true probability that the i-th image belongs to the j-th category, p′ ij represents the predicted probability that the i-th image belongs to the j-th class, and N represents the total number of samples in the data set, j represents the number of samples in the jth class, w j Expressed as the class weight of the jth class;

[0061] The image decomposition model adopts a convolutional neural network. The image decomposition model includes a convolution layer and a merging layer. The first layer of the image decomposition model includes three convolution layers. The number of neurons in the three convolution layers is 12, 20, and 32 respectively. The convolution kernel sizes of the three convolution layers are 3*3, 5*5, and 7*7 respectively. Different image features are extracted through the first layer. The second layer is the merging layer. Each feature map is split according to the channel dimension through the second layer and processed separately. The processed feature maps are spliced ​​on the channel dimension to form a new feature map. The third and fourth layers are both convolution layers. The number of neurons is 64 and the convolution kernel size is 3*3. The image features are refined through the third and fourth layers. The fifth layer is a convolution layer. The number of neurons is 4 and the convolution kernel size is 3*3. The image features are decomposed into reflectance map and illumination map through the fifth layer, and the reflectance map and illumination map are output. Finally, the fifth layer sets the Sigmoid activation function. The pixel values ​​of the output reflectance map and illumination map are normalized to [0,1] through the Sigmoid activation function.

[0062] The loss function of the image decomposition model is:

[0063] L=L s +λ r L r +λ l L l

[0064]

[0065] L r =||R low -R normal ||2

[0066]

[0067] Among them, L s Represents the loss function of the reconstructed image after image restoration, L r Represents the loss function for the reflectivity map to remain unchanged, L l represents the loss function of the smooth illumination map, λ r and λ l The constants representing the constancy of the balanced reflectance map and the smoothness of the illumination map, R i represents the normal lighting image vector of the i-th category, S j represents the j-th normal lighting image vector, L i represents the light map vector, represents smoothness constraint, R normal Represents the reflectivity map, R low Reflectance map representing low-light images;

[0068] The reflectance map refinement model adopts a convolutional neural network. The reflectance map refinement model includes convolution layers. The first layer of the reflectance map refinement model is a convolution layer with 64 neurons, a convolution kernel of 3*3, and an activation function of PReLu. The coarsely extracted reflectance map is convolved through the first layer. The second layer is a convolution layer with 64 neurons, a convolution kernel of 3*3, and an activation function of PReLu. The reflectance map after the convolution of the first layer is reconstructed through the second layer. The third layer is a convolution layer with 3 neurons, a convolution kernel of 3*3, and an activation function of Sigmoid. The reflectance map reconstructed in the second layer is restored in detail through the third layer, and the pixel value of the output reflectance map is normalized to [0,1] through the Sigmoid activation function.

[0069] The loss function of the reflectance map refinement model is:

[0070] L=||P′-P||2

[0071] P=R normal -R low

[0072] Where P′ represents R normal and R low The residual, P represents the true residual, R normal Represents the reflectivity map, R low Reflectance map representing low-light images;

[0073] The illumination map enhancement model adopts a convolutional neural network. The illumination map enhancement model includes convolutional layers and connection layers. The first three layers of the reflectance map refinement model are all convolutional layers, with 64 neurons, 3*3 convolution kernels, and PReLu activation functions. The first three layers are encoded and downsampled to extract global illumination map features. The middle three layers are all convolutional layers, with 64 neurons, 3*3 convolution kernels, and PReLu activation functions. The enhanced features of the illumination map are obtained through the middle three layers. The seventh layer is the connection layer, and skip connections are performed through the seventh layer to combine multi-layer features. The eighth layer is a convolutional layer with 1 neuron, 3*3 convolution kernels, and Sigmoid activation functions. The pixel values ​​of the output enhanced illumination map are normalized to [0,1] through the Sigmoid activation function.

[0074] The loss function of the light map enhancement model is:

[0075] L=L1+λ1L2+λ2L3

[0076]

[0077] L3=||L′ low -L normal ||2

[0078] Among them, L represents the total loss function, L1 represents the reconstruction loss function, L2 represents the loss function based on the characteristics of the illumination map, L3 represents the consistency loss function between the enhanced illumination map and the normal illumination image, and L′ low represents the enhanced illumination map, S normal represents the normal lighting image, λ r and λ l constants representing the constancy of the light map properties and the consistency of the light map, respectively;

[0079] In S7, the refined reflectance map and the enhanced illumination map are subjected to gradient weighting to obtain an enhanced image, including the following steps:

[0080] S701, taking the refined reflectivity map and the enhanced illumination map as input;

[0081] S702, performing gradient calculation on the reflectivity map and the illumination map, obtaining the gradient magnitude and direction of the reflectivity map and the illumination map, and multiplying the gradient magnitude by a weight coefficient;

[0082] S703, weighting the two gradient images according to the magnitude and direction of the gradient amplitude;

[0083] S704: Fusing the weighted gradient image with the original low-light image using a weighted superposition method to obtain an enhanced image.

[0084] Please refer to Figure 3 , the normal illumination image is input into the image reconstruction model, the output end of the image reconstruction model outputs a low illumination image, and the low illumination image is input into the image detection model, the image detection model detects whether the low illumination image is qualified, the qualified low illumination image is input into the image decomposition model, the output end of the image decomposition model outputs a reflection map and an illumination map, and the reflection map is input into the reflection map refinement model, the reflection map refinement model outputs a refined reflection map, the illumination map is input into the illumination map enhancement model, the illumination map enhancement model outputs an enhanced illumination map, the refined reflection map and the enhanced illumination map are gradient weighted to form an enhanced image;

[0085] Please refer to Figure 4 , detected as a low-light image by the image detection model;

[0086] See also Figure 5 , the low illumination image is input to the image enhancement module, and the enhanced image is output through the image enhancement module.

[0087] It should be noted that, in this document, relational terms such as first and second, etc., are used only 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 "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0088] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A low-light image enhancement method based on convolutional neural network, characterized in that: The following steps are involved: S1. Obtain a normal illumination image, use an image reconstruction model to artificially synthesize a low illumination image, and use the synthesized low illumination image as an image set; S2. Divide the acquired image set into a first training set and a first test set according to corresponding proportions, and train the image detection model using the first training set to obtain a trained image detection model, and perform a performance test on the trained image detection model using the first test set, where the input end of the image detection model is the image, and the output end of the image detection model is the image detected as low-light; S3. Dividing the images detected as low-illuminance into a second training set and a second test set, and training the image enhancement model using the second training set to obtain a trained image enhancement model, and performing a performance test on the trained image enhancement model using the second test set, where the input end of the image enhancement model is the low-illuminance image, and the output end of the image enhancement model is the enhanced image; S4. The image enhancement model includes an image decomposition model, a reflectance map refinement model, and an illumination map enhancement model. The image decomposition model is trained using the second training set to obtain a trained image decomposition model. The input of the image decomposition model is the low-light image, and the output of the image decomposition model is the reflectance map and the illumination map. S5. Training a reflectivity map refinement model using the reflectivity map to obtain a trained reflectivity map refinement model, wherein an input end of the reflectivity map refinement model is the reflectivity map, and an output end of the reflectivity map refinement model is the refined reflectivity map; S6. Training a light map enhancement model using the light map to obtain a trained light map enhancement model, wherein an input end of the light map enhancement model is the light map, and an output end of the light map enhancement model is the enhanced light map; S7, performing gradient weighting on the refined reflectance map and the enhanced illumination map to obtain an enhanced image; The image detection model adopts a convolutional neural network, which includes a convolution layer, a pooling layer and a fully connected layer. The convolution layer, the pooling layer and the fully connected layer all have 3 layers. The first layer of the image detection model is a convolution layer, the number of neurons is 64, the convolution kernel size is 7*7, and the step size is 1. The second layer is a pooling layer, the convolution kernel size is 2*2, and the step size is 2. The third layer is a convolution layer, the number of neurons is 128, the convolution kernel size is 5*5, and the step size is 1. The fourth layer is a pooling layer, the convolution kernel size is 2*2, and the step size is 2. The fifth layer is a convolutional layer with 256 neurons, a convolution kernel size of 3*3, and a step size of 1. The sixth layer is a pooling layer with a convolution kernel size of 2*2 and a step size of 2. The seventh and eighth layers are fully connected layers with 512 neurons each. The ninth layer is a fully connected layer with 2 neurons. The first, third, fifth, seventh, and eighth layers are all provided with activation functions, and the activation functions are all relu functions. The first six layers of the image detection model are used to extract image features, and the last three layers of the image detection model are used for image classification. The image decomposition model adopts a convolutional neural network. The image decomposition model includes a convolution layer and a merging layer. The first layer of the image decomposition model includes three convolution layers. The number of neurons in the three convolution layers is 12, 20 and 32 respectively. The convolution kernel sizes of the three convolution layers are 3*3, 5*5 and 7*7 respectively. Different image features are extracted through the first layer. The second layer is a merging layer. Each feature map is split according to the channel dimension through the second layer and processed separately. The processed feature maps are spliced ​​in the channel dimension to form a new Feature map, the third and fourth layers are both convolution layers, the number of neurons is 64, the convolution kernel size is 3*3, the image features are refined through the third and fourth layers, the fifth layer is a convolution layer, the number of neurons is 4, the convolution kernel size is 3*3, the image features are decomposed into reflectance map and illumination map through the fifth layer, and the reflectance map and illumination map are output. Finally, the fifth layer sets the Sigmoid activation function, and the pixel values ​​of the output reflectance map and illumination map are normalized to [0,1] through the Sigmoid activation function; The reflectance map refinement model adopts a convolutional neural network, which includes a convolution layer. The first layer of the reflectance map refinement model is a convolution layer, with 64 neurons, a convolution kernel of 3*3, and an activation function of PReLu function. The coarsely extracted reflectance map is convolved through the first layer. The second layer is a convolution layer, with 64 neurons, a convolution kernel of 3*3, and an activation function of PReLu function. The reflectance map after the convolution of the first layer is reconstructed through the second layer. The third layer is a convolution layer, with 3 neurons, a convolution kernel of 3*3, and an activation function of Sigmoid function. The reflectance map reconstructed in the second layer is restored in detail through the third layer, and the pixel value of the output reflectance map is normalized to [0,1] through the Sigmoid activation function. The loss function of the reflection map refinement model is: ； ； in, express and The residual, represents its true residual, represents the reflectivity map, Reflectance map representing low-light images; In S7, the enhanced image is obtained by performing gradient weighting on the refined reflectance map and the enhanced illumination map, including the following steps: S701, taking the refined reflectivity map and the enhanced illumination map as input; S702, performing gradient calculation on the reflectivity map and the illumination map, obtaining the gradient magnitude and direction of the reflectivity map and the illumination map, and multiplying the gradient magnitude by a weight coefficient; S703, weighting the two gradient images according to the magnitude and direction of the gradient amplitude; S704: Fusing the weighted gradient image with the original low-light image using a weighted superposition method to obtain an enhanced image.

2. The low-light image enhancement method based on convolutional neural network according to claim 1, characterized in that: The loss function of the image detection model is: ； in, Indicates the number of images, Indicates the number of categories of images, Expressed as The class weight of the class, Indicates the The image belongs to The true probability of the class, Indicates the Class images belong to The predicted probability of the class, and , Indicates the total number of samples in the dataset, Indicates the The number of class samples, It is expressed as the class weight of the jth class.

3. The low-light image enhancement method based on convolutional neural network according to claim 1, characterized in that: The light map enhancement model adopts a convolutional neural network. The light map enhancement model includes a convolution layer and a connection layer. The first three layers of the light map enhancement model are all convolution layers, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is PReLu function. The first three layers are encoded and downsampled to extract the global light map features. The middle three layers are all convolution layers, the number of neurons is 64, the convolution kernel is 3*3, and the activation function is PReLu function. The enhanced features of the light map are obtained through the middle three layers. The seventh layer is a connection layer, and skip connection is performed through the seventh layer to combine the multi-layer features. The eighth layer is a convolution layer, the number of neurons is 1, the convolution kernel is 3*3, and the activation function is Sigmoid function. The pixel value of the output enhanced light map is normalized to [0,1] through the Sigmoid activation function.

Images