A Low-Light Image Enhancement Method Based on Spectral Interaction

Abstract

This invention belongs to the field of low-light image enhancement, specifically relating to a low-light image enhancement method based on spectral interaction. The method includes constructing a low-light image enhancement model, which comprises a spatial feature extraction module, a spectral real part interaction module, and a brightness-perceived fusion module. The model is trained, and the trained low-light image enhancement model is used to enhance low-light images. This invention employs a two-branch fusion strategy, which effectively utilizes the advantages of both branches while avoiding their respective shortcomings, further improving image quality and clarity.

Description

Technical Field

[0001] This invention belongs to the field of low-light image enhancement, and specifically relates to a low-light image enhancement method based on spectral interaction. Background Technology

[0002] In real life, low-light images are ubiquitous due to environmental issues or technological limitations, such as object occlusion and restricted shooting conditions. Low-light images are typically of poor quality, resulting in poor visualization and impacting subsequent downstream tasks such as semantic segmentation, face detection, and autonomous driving. Therefore, low-light image enhancement has been proposed to restore areas of degraded or lost illumination in images while suppressing noise interference, thereby significantly improving image quality and visibility.

[0003] In recent years, many methods have been proposed to address low-light image enhancement. These methods can be broadly categorized into two types: traditional methods (such as histogram equalization, gamma correction, and Retinex-based methods) and deep learning-based algorithms. The former has relatively low computational complexity and therefore faster processing speed, but its enhancement effect is limited. It fails to achieve the desired results for images under certain extreme low-light conditions and may amplify noise in the image, leading to a decrease in the quality of the enhanced image. The latter can achieve better enhancement results by training models on large datasets to optimize for specific scenes or applications. However, this requires significant computational resources and time to train the model and has limited generalization ability.

[0004] Frequency domain methods are an important component of image processing. They extract frequency domain features from images and are used in computational vision tasks such as image classification, object detection, and facial super-resolution, all of which have proven to be effective. Recently, Yin et al. used discrete cosine transform (DCT) to segment the frequency domain features of images and recover image details. However, DCT requires significant computational resources and time, and its computational complexity may become a bottleneck, especially when processing large-scale data. Li et al. explored the feasibility of Fast Fourier Transform (FFT) in low-light image enhancement, but only observed that in the frequency domain, brightness information is contained in the amplitude component, and noise is contained in the phase component. They did not consider that the difference between the amplitude component of images with different brightness levels and the real image is significant, and that the phase component also contains important image information.

[0005] In summary, existing technologies, when transforming low-light images from the spatial domain to the frequency domain, often rely on stacked network structures and only consider the effect of the amplitude component, while neglecting the importance of the phase component in image restoration. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a low-light image enhancement method based on spectral interaction. The method includes constructing a low-light image enhancement model, which comprises a spatial feature extraction module, a spectral real part interaction module, and a brightness-perceived fusion module. The model is trained, and the trained low-light image enhancement model is used to enhance low-light images. The training process of the low-light image enhancement model includes:

[0007] Construct a training dataset. The training data in the training dataset consists of sample pairs. Each sample pair includes a low-light image and a normal-light image of the same scene.

[0008] The low-light image in the sample pair is inverted to obtain its overexposed image. The spatial feature extraction module is used to extract the spatial feature maps of the low-light image and its overexposed image respectively.

[0009] The real part of the spectrum interaction module transforms the spatial domain to the frequency domain based on the fast Fourier transform, thereby obtaining the frequency domain feature maps of the low-light image and its overexposed image.

[0010] The frequency domain feature maps of the low-light image and its overexposed image are fed into the brightness-sensing fusion module for fusion, and the enhanced image is reconstructed.

[0011] A loss function for the low-light image enhancement model is constructed based on the reconstruction loss, frequency domain loss, and perceptual loss of the enhanced image and the normal-light image. The low-light image enhancement model is then optimized based on the loss function.

[0012] Furthermore, the process of inverting a low-light image to obtain its overexposed image includes:

[0013]

[0014] in, Represents a low-light image; Representing an image Overexposed images; These are learnable parameters.

[0015] Furthermore, the spatial feature extraction module includes 1×3 convolutional units, 3×3 convolutional units, 3×1 convolutional units, a channel stitching unit, and a curve adjustment unit. The image is processed by the 1×3, 3×3, and 3×1 convolutional units respectively to obtain feature maps of three different sizes. These three feature maps are then stitched together via channels. The curve adjustment unit performs curve adjustment on the stitched image. The adjustment process includes:

[0016]

[0017] in, The feature map is obtained after curve adjustment of an image of type i. When i=1, it means that the image type is a low-light image, and when i=2, it means that the image type is an overexposed image. A multi-size stitched image representing an image of type i; These are the parameters of the learnable curve.

[0018] Furthermore, the process by which the spectrum real part interaction module obtains the frequency domain feature maps of the low-light image and its overexposed image includes:

[0019] Fast Fourier Transform is performed on the spatial feature maps of the low-light image and its overexposed image, respectively, and the amplitude and phase representations of the image are obtained based on the transformation results.

[0020] By combining the amplitude and phase using Euler's formula, the real and imaginary parts of the image can be obtained;

[0021] Swap the real parts of the low-light image and its overexposed image, and perform an inverse fast Fourier transform on the swapped image to obtain the frequency domain feature maps of the low-light image and its overexposed image.

[0022] Furthermore, the process of feeding the frequency domain feature maps of the low-light image and its overexposed image into the brightness-sensing fusion module for fusion includes:

[0023] The frequency domain feature maps of the low-light image are subjected to global max pooling and global average pooling respectively. The shape of the pooled image is adjusted so that the dimension of the pooled image is [B,N,H×W]. The two adjusted images are multiplied at the pixel level to obtain the brightness information of the low-light image.

[0024] Global max pooling is performed on the frequency domain feature map of the overexposed image. The shape of the pooled image is adjusted so that the dimensions of the pooled image are [B,N,H×W]. After adjusting the shape, the brightness information of the overexposed image is obtained.

[0025] The enhanced image is obtained by multiplying the brightness information of the low-light image and the exposed image at the pixel level and then performing a 3×3 convolution.

[0026] Where B is the number of images input to the low-light image enhancement model in one operation; N is the number of channels in the image; H is the height of the image; and W is the width of the image.

[0027] Furthermore, the loss function of the low-light image enhancement model is expressed as:

[0028]

[0029] in, The loss function for the low-light image enhancement model; The reconstruction loss for low-light image enhancement models; Frequency domain loss for low-light image enhancement models; To enhance the perceptual loss of the model for low-light images; , To balance and Hyperparameters.

[0030] Furthermore, the reconstruction loss of the low-light image enhancement model Represented as:

[0031]

[0032]

[0033] in, Represents the reconstructed image Compared with normal lighting images Similarity function between them; express Loss, specifically absolute loss; To reconstruct the image The average value of a pixel; Normal lighting image The average value of a pixel; To reconstruct the image Pixel variance; Normal lighting image Pixel variance; Represents the reconstructed image Compared with normal lighting images Pixel covariance; and These are two constants.

[0034] Furthermore, the frequency domain loss of the low-light image enhancement model Represented as:

[0035]

[0036] in, express Loss, specifically absolute loss; Indicates the reconstructed image Perform a Fast Fourier Transform. This indicates that the amplitude components are calculated based on the results of the Fast Fourier Transform. Indicates normal lighting image Perform a fast Fourier transform; This indicates that the phase component is calculated based on the results of the Fast Fourier Transform.

[0037] Furthermore, the process of performing a Fast Fourier Transform on an image includes:

[0038]

[0039] in, This represents the coordinates of image x in space at a height of h and a width of w, where H is the height of the image and W is the width of the image; The coordinates of the image x in Fourier space are... The value at;

[0040]

[0041]

[0042] in, Indicates taking The real part, express The square of; Indicates taking The imaginary part, express The square of.

[0043] Perceptual loss of low-light image enhancement models Represented as:

[0044]

[0045] in, express Loss, specifically absolute loss; The term "feature extraction network" refers to the VGG network used in this invention. Those skilled in the art can choose a feature extraction network based on actual application requirements.

[0046] This invention designs a low-light image enhancement model based on Fast Fourier Transform (FFT). In this model, an overexposed image is obtained through an image inversion mechanism, constructing a dual-branch structure to address the problem of limited training data. This step effectively utilizes the complementarity between the overexposed image and the original low-light image, expanding the training dataset and improving the model's generalization ability. Next, a multi-size convolutional kernel strategy is used to extract image features, effectively enhancing local image information, increasing network nonlinearity, and reducing parameters and computational cost. This strategy can better capture image features at different spatial scales, improving the model's ability to perceive and understand image details. Then, the image is transformed from the spatial domain to the frequency domain for processing, handling the amplitude and phase components separately, and obtaining the real and imaginary parts using Euler's formula. By transforming the image from the spatial domain to the frequency domain, the frequency information of the image can be better analyzed and utilized, further enhancing the image quality and clarity. Next, interacting the real parts of the two branches allows for full utilization of complementary information to recover the image. This step, by fusing the real parts of different branches, better utilizes the complementarity between the different branches of the model, improving the model's recovery effect. Finally, the images of the two branches are organically fused using a brightness-aware strategy. This step, through the organic fusion of the images of the two branches using a brightness-aware strategy, can more comprehensively and accurately recover and enhance the image information. In summary, the two-branch fusion strategy of this invention can effectively utilize the advantages of both branches, avoid their respective shortcomings, and further improve the image quality and clarity. Attached Figure Description

[0047] Figure 1 This is a flowchart of a low-light image enhancement method based on spectrum interaction according to the present invention;

[0048] Figure 2 This is a flowchart of the spatial feature extraction module of the present invention;

[0049] Figure 3 This is a flowchart of the spectrum real part interaction module of the present invention;

[0050] Figure 4 This is a flowchart of the brightness perception fusion module of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] This invention proposes a low-light image enhancement method based on spectral interaction, comprising constructing a low-light image enhancement model, which includes a spatial feature extraction module, a spectral real part interaction module, and a brightness-perceived fusion module; training the model; and using the trained low-light image enhancement model to enhance low-light images. The training process of the low-light image enhancement model includes:

[0053] Construct a training dataset. The training data in the training dataset consists of sample pairs. Each sample pair includes a low-light image and a normal-light image of the same scene.

[0054] The low-light image in the sample pair is inverted to obtain its overexposed image. The spatial feature extraction module is used to extract the spatial feature maps of the low-light image and its overexposed image respectively.

[0055] The real part of the spectrum interaction module transforms the spatial domain to the frequency domain based on the fast Fourier transform, thereby obtaining the frequency domain feature maps of the low-light image and its overexposed image.

[0056] The frequency domain feature maps of the low-light image and its overexposed image are fed into the brightness-sensing fusion module for fusion, and the enhanced image is reconstructed.

[0057] A loss function for the low-light image enhancement model is constructed based on the reconstruction loss, frequency domain loss, and perceptual loss of the enhanced image and the normal-light image. The low-light image enhancement model is then optimized based on the loss function.

[0058] In this embodiment, when implementing a low-light image enhancement method based on spectral interaction, such as... Figure 1 As shown, it includes the following steps:

[0059] S1: Construct a paired dataset consisting of low-light images and normal-light images, wherein each low-light image corresponds to a normal-light image of the same scene, for training the low-light image enhancement model.

[0060] S2: Construct a low-light image enhancement model; the model includes a spatial feature extraction module, a spectrum real part interaction module, and a brightness perception fusion module;

[0061] S3: Input the low-light images from the training dataset into the low-light image enhancement model and train the model. This includes the following steps:

[0062] S31: Input the low-light image I1, and obtain the overexposed image I2 that is paired with it through the inversion mechanism to construct a dual-branch network structure.

[0063] S32: The two obtained images are input into the spatial domain feature extraction module to extract local information from the images. The spatial domain feature extraction module consists of multi-size convolutional kernels, including 1×3 convolutional units, 3×3 convolutional units, 3×1 convolutional units, channel stitching units, and curve adjustment units. The two images output by the spatial domain feature extraction module are denoted as follows: , ;

[0064] S33: Input the feature-extracted image into the real part of the spectrum interaction module to extract the frequency characteristics of the image. The image is transformed from the spatial domain to the frequency domain using a Fast Fourier Transform. The two images output by the real part of the spectrum interaction module are denoted as follows: , ;

[0065] S34: The image after frequency domain processing is sent to the brightness perception fusion module to obtain complementary information of the left and right branches;

[0066] S35: Reconstruction loss, frequency domain loss, and perceptual loss are used as the final loss function of the low-light image enhancement model. During model training, the model is trained by minimizing the loss function.

[0067] S4: Inputting the low-light image to be detected into the already trained model can yield enhanced results.

[0068] In this embodiment, the problem of insufficient training data is solved by using an inversion mechanism. The inversion mechanism can be expressed as follows:

[0069]

[0070] in, This indicates that the input image is a low-light image; This represents the overexposed image paired with the input image; It is a learnable parameter used to adjust the exposure of an image.

[0071] In this embodiment, a multi-size convolution strategy is employed to leverage the dependencies between adjacent image patches and to enhance the correlation between channels by adjusting curves, including:

[0072] In the spatial domain feature extraction module, two images obtained through an inversion mechanism are input, and features from different receptive fields are extracted using 1×3, 3×3, and 3×1 convolutions. This enhances the local information of the image to varying degrees, increases the nonlinearity of the network, and reduces parameters and computational cost. Subsequently, an adjustment curve is used to better preserve the inherent colors of the image and reduce the risk of oversaturation, which can be expressed as:

[0073]

[0074] in, The feature map is obtained after curve adjustment of an image of type i. When i=1, it means that the image type is a low-light image, and when i=2, it means that the image type is an overexposed image. A multi-size stitched image representing an image of type i; These are the parameters of the learnable curve.

[0075] Obtaining the frequency characteristics of an image by transforming it from the spatial domain to the frequency domain includes:

[0076] For Fourier transform, given an input image Its size is × The Fourier transform is expressed as:

[0077]

[0078] in, denoted by Fourier transform; h and w are the coordinates of image x in space, u and v represent the coordinates in Fourier space, and j represents the imaginary unit; This represents the coordinates of image x in space at a height of h and a width of w. The coordinates of the image x in Fourier space are... The value at that location. The above formula can be simplified to:

[0079]

[0080] in, Indicates taking The real part, Indicates taking The imaginary part.

[0081] After the above Fast Fourier Transform, the real parts of the low-light image and its overexposed image can be swapped, allowing the information of the low-light image and its overexposed image to interact. Then, the inverse Fast Fourier Transform is directly performed to obtain the frequency domain feature maps of the low-light image and its overexposed image.

[0082] In this embodiment, a method for obtaining the real and imaginary parts of an image is also provided. After a Fast Fourier Transform, the amplitude and phase components of the image can be calculated. The amplitude component contains a large amount of brightness information, while the phase component contains noise and other information. In the frequency domain, the two effective components of the image x, namely the amplitude and phase components, can be obtained by the following formula:

[0083]

[0084]

[0085] in, express The square of, express The square of.

[0086] Transforming the image from the spatial domain to the frequency domain after the spatial feature extraction module yields the amplitude and phase components of the left and right branches. Then, Euler's formula is used to combine the amplitude and phase components, obtaining their corresponding real and imaginary parts. Interacting the real parts of the left and right branches utilizes information from both the amplitude and phase components, enabling a more comprehensive analysis of image features. This process also avoids losing details, allowing the model to better understand and process the image. The operation of Euler's formula can be written as:

[0087]

[0088]

[0089] in, and These represent the amplitude component and phase component of the image of type i, respectively; Represents the real part of an image of type i; Let i represent the imaginary part of the image of type i. Performing an inverse fast Fourier transform on the swapped images yields the frequency domain feature maps of the low-light image and its overexposed image. The spectral interaction process can be represented as:

[0090]

[0091]

[0092] in, This represents the frequency domain feature map obtained from the low-light image; This represents the inverse fast Fourier transform; This represents the real part of the overexposed image after processing with Euler's formula; This represents the imaginary part of a low-light image after processing with Euler's formula. This represents the frequency domain feature map obtained from the overexposed image; This represents the real part of the low-light image after processing with Euler's formula; This represents the imaginary part of the overexposed image after processing with Euler's formula.

[0093] In this embodiment, considering the insufficient correlation between the left and right branches, the brightness-aware fusion module can effectively enhance the correlation between the left and right branches and improve the performance of the model. The brightness-aware strategy to capture the complementary information of the left and right branches specifically includes the following steps:

[0094] For the left branch, the right branch is the processing of overexposed images, which naturally contains a large amount of brightness information. This can help the left branch better restore the brightness of the image. That is, low-light images are easier to capture details and color information. Therefore, reasonable fusion can obtain a more clear contrast, rich detail information, and enhancement results without overexposure. Therefore, global max pooling and global average pooling are performed on the frequency domain feature maps of the low-light image respectively. After adjusting the shape of the pooled image, the two pooled feature maps are multiplied at the pixel level to obtain the brightness information of the low-light image. In this embodiment, the dimension of the low-light image is [B,N,H,W]. The pooling in this invention does not compress the image, and the dimension after pooling is [B,N,H,W]. The shape of the pooled image is adjusted, and the dimension of the pooled image [B,N,H,W] is adjusted to [B,N,H×W], where B is the number of images input to the low-light image enhancement model at one time; N is the number of channels of the image; H is the height of the image; and W is the width of the image.

[0095] Global max pooling is performed on the frequency domain feature map of the overexposed image. The brightness information of the overexposed image is obtained by adjusting the shape of the pooled image. The shape adjustment of the image is the same as that of the low-light image, that is, the dimension [B,N,H,W] of the pooled image is adjusted to [B,N,H×W].

[0096] The enhanced image is obtained by multiplying the brightness information of the low-light image and the exposed image at the pixel level and then performing a 3×3 convolution.

[0097] In this embodiment, a loss function for the low-light image enhancement model is constructed based on the reconstruction loss, frequency domain loss, and perceptual loss of the enhanced image and the normal-light image. The low-light image enhancement model is then optimized based on this loss function. The loss function of the low-light image enhancement model is expressed as follows:

[0098]

[0099] in, The loss function for the low-light image enhancement model; The reconstruction loss for low-light image enhancement models; Frequency domain loss for low-light image enhancement models; To enhance the perceptual loss of the model for low-light images; , To balance and Hyperparameters.

[0100] In this embodiment, the reconstruction loss of the low-light image enhancement model Represented as:

[0101]

[0102]

[0103] in, Represents the reconstructed image Compared with normal lighting images Similarity function between them; express Loss, specifically absolute loss; To reconstruct the image The average value of a pixel; Normal lighting image The average value of a pixel; To reconstruct the image Pixel variance; Normal lighting image Pixel variance; Represents the reconstructed image Compared with normal lighting images Pixel covariance; and These are two constants, parameters set to prevent the denominator from being zero.

[0104] In this embodiment, the frequency domain loss of the low-light image enhancement model Represented as:

[0105]

[0106] in, Indicates the reconstructed image Perform a Fast Fourier Transform. This indicates that the amplitude components are calculated based on the results of the Fast Fourier Transform. Indicates normal lighting image Perform a fast Fourier transform; This indicates that the phase component is calculated based on the results of the Fast Fourier Transform.

[0107] In this embodiment, the perceptual loss of the low-light image enhancement model Represented as:

[0108]

[0109] in, express Loss, specifically absolute loss; Indicates from The operation of extracting features from a network.

[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-light image enhancement method based on spectral interaction, characterized in that, A low-light image enhancement model is constructed, comprising a spatial feature extraction module, a spectrum real part interaction module, and a brightness-perceived fusion module. This model is trained, and the trained low-light image enhancement model is then used to enhance low-light images. The training process for the low-light image enhancement model includes: Construct a training dataset. The training data in the training dataset consists of sample pairs. Each sample pair includes a low-light image and a normal-light image of the same scene. The low-light image in the sample pair is inverted to obtain its overexposed image. The spatial feature extraction module is used to extract the spatial feature maps of the low-light image and its overexposed image respectively. The real part of the spectrum interaction module transforms the spatial domain to the frequency domain based on the fast Fourier transform, thereby obtaining the frequency domain feature maps of the low-light image and its overexposed image. The frequency domain feature maps of the low-light image and its overexposed image are fed into the brightness-sensing fusion module for fusion, and the enhanced image is reconstructed. A loss function for the low-light image enhancement model is constructed based on the reconstruction loss, frequency domain loss, and perceptual loss of the enhanced image and the normal-light image. The low-light image enhancement model is then optimized based on the loss function.

2. The low-light image enhancement method based on spectral interaction according to claim 1, characterized in that, The process of inverting a low-light image to obtain its overexposed image includes: in, Represents a low-light image; Representing an image Overexposed images; These are learnable parameters.

3. The low-light image enhancement method based on spectral interaction according to claim 1, characterized in that, The spatial feature extraction module includes 1×3 convolutional units, 3×3 convolutional units, 3×1 convolutional units, a channel stitching unit, and a curve adjustment unit. The image is processed by the 1×3, 3×3, and 3×1 convolutional units respectively to obtain feature maps of three different sizes. These three feature maps are then stitched together via channels. The curve adjustment unit performs curve adjustment on the stitched image. The adjustment process includes: in, The feature map is obtained after curve adjustment of an image of type i. When i=1, it means that the image type is a low-light image, and when i=2, it means that the image type is an overexposed image. A multi-size stitched image representing an image of type i; These are the parameters of the learnable curve.

4. The low-light image enhancement method based on spectral interaction according to claim 1, characterized in that, The process by which the real part of the spectrum interaction module obtains the frequency domain feature maps of the low-light image and its overexposed image includes: Fast Fourier Transform is performed on the spatial feature maps of the low-light image and its overexposed image, respectively, and the amplitude and phase representations of the image are obtained based on the transformation results. By combining the amplitude and phase using Euler's formula, the real and imaginary parts of the image can be obtained; Swap the real parts of the low-light image and its overexposed image, and perform an inverse fast Fourier transform on the swapped image to obtain the frequency domain feature maps of the low-light image and its overexposed image.

5. The low-light image enhancement method based on spectral interaction according to claim 1, characterized in that, The process of fusing the frequency domain feature maps of the low-light image and its overexposed image into the brightness-sensing fusion module includes: The frequency domain feature maps of the low-light image are subjected to global max pooling and global average pooling respectively. The shapes of the two pooled images are adjusted so that the dimensions of the pooled image are [B,N,H×W]. The two adjusted images are multiplied at the pixel level to obtain the brightness information of the low-light image. Global max pooling is performed on the frequency domain feature map of the overexposed image. The shape of the pooled image is adjusted so that the dimensions of the pooled image are [B,N,H×W]. After adjusting the shape, the brightness information of the overexposed image is obtained. The enhanced image is obtained by multiplying the brightness information of the low-light image and the exposed image at the pixel level and then performing a 3×3 convolution. Where B is the number of images input to the low-light image enhancement model in one operation; N is the number of channels in the image; H is the height of the image; and W is the width of the image.

6. The low-light image enhancement method based on spectral interaction according to claim 1, characterized in that, The loss function of the low-light image enhancement model is expressed as: in, The loss function for the low-light image enhancement model; The reconstruction loss for low-light image enhancement models; Frequency domain loss for low-light image enhancement models; To enhance the perceptual loss of the model for low-light images; , To balance and Hyperparameters.

7. The low-light image enhancement method based on spectral interaction according to claim 6, characterized in that, Reconstruction loss of low-light image enhancement models Represented as: in, Represents the reconstructed image Compared with normal lighting images Similarity function between them; express Loss, specifically absolute loss; To reconstruct the image The average value of a pixel; Normal lighting image The average value of a pixel; To reconstruct the image Pixel variance; Normal lighting image Pixel variance; Represents the reconstructed image Compared with normal lighting images Pixel covariance; and These are two constants.

8. The low-light image enhancement method based on spectral interaction according to claim 6, characterized in that, Frequency domain loss of low-light image enhancement model Represented as: in, express Loss, specifically absolute loss; Indicates the reconstructed image Perform a Fast Fourier Transform. This indicates that the amplitude components are calculated based on the results of the Fast Fourier Transform. Indicates normal lighting image Perform a fast Fourier transform; This indicates that the phase component is calculated based on the results of the Fast Fourier Transform.

9. A low-light image enhancement method based on spectral interaction according to claim 4 or 8, characterized in that, The process of performing a Fast Fourier Transform on an image includes: in, This represents the coordinates of image x in space at a height of h and a width of w, where H is the height of the image and W is the width of the image; The coordinates of the image x in Fourier space are... The value at; in, Indicates taking The real part, express The square of; Indicates taking The imaginary part, express The square of.

10. A low-light image enhancement method based on spectral interaction according to claim 6, characterized in that, Perceptual loss of low-light image enhancement models Represented as: in, express Loss, specifically absolute loss; This represents a feature extraction network.

Images