Low-illumination image enhancement method, electronic equipment and storage medium

Abstract

The invention discloses a low-illumination image enhancement method, electronic equipment and a storage medium, and relates to the field of digital image processing. The low-illumination image enhancement method comprises the following steps of S1, performing discrete wavelet transform decomposition on a low-illumination image and a normal-illumination image based on multiple discrete wavelets to obtain sub-bands; s2, building a multi-scale vision-language guidance network, performing appearance guidance in combination with a frequency domain diffusion and multi-modal method, and performing diffusion on a low-frequency domain sub-band {AK} by using a vision-language prompt and diffusion model to realize low-frequency domain enhancement of the low-illumination image; s3, the high-frequency sensing module processes the high-frequency domain sub-band based on Fourier transform to realize high-frequency domain enhancement of the low-illumination image; and S4, performing inverse wavelet transform on the enhanced low-frequency sub-band and high-frequency sub-band to obtain a final enhanced image. The image is decomposed into the low-frequency domain and the high-frequency domain, so that the spatial dimension is greatly reduced, and the calculation complexity and the resource consumption are reduced.

Description

Technical Field

[0001] The present application relates to the field of digital image processing, and in particular to a low-light image enhancement method, electronic device, and storage medium. Background Art

[0002] Currently, due to complex lighting conditions, captured images may lose relevant information due to insufficient illumination. This phenomenon poses a challenge to human visual perception and hinders the development and deployment of downstream tasks such as object detection, autonomous driving, and text detection. To address these issues, low-light image enhancement technology has rapidly developed.

[0003] However, existing methods still lack stability and visual perceptual quality in image restoration. Traditional low-light image enhancement methods rely primarily on single-modality image-level supervision, which leads to difficulties in content reconstruction and reduced visual quality. Furthermore, while diffusion models have strong generative capabilities, their diversity becomes a limiting factor in image enhancement and restoration tasks. Summary of the Invention

[0004] To solve the above problems, the present application discloses a low-light image enhancement method, which significantly improves the quality and visual perception effect of low-light images by combining frequency domain diffusion, multimodal guidance and high-frequency information recovery.

[0005] The first technical solution adopted in this application is to provide a low-light image enhancement method, comprising the following steps: S1: Based on multiple discrete wavelets, the low-light image and the normal-light image are decomposed into discrete wavelet transforms to obtain subbands: {AK, VK, HK, DK}; where AK represents the low-frequency domain, VK, HK, and DK represent the high-frequency domains in the vertical, horizontal, and diagonal directions, respectively; S2: Build a multi-scale visual-linguistic guidance network, combine frequency-domain diffusion and multimodal methods for appearance guidance, and use visual-linguistic cues and diffusion models to diffuse the low-frequency subband {AK} to achieve low-frequency domain enhancement of low-light images; S3: The high-frequency perception module processes the high-frequency subbands {VK, HK, DK} based on Fourier transform to achieve high-frequency domain enhancement of low-light images; S4: The enhanced low-frequency sub-band and high-frequency sub-band are subjected to inverse wavelet transform to obtain the final enhanced image.

[0006] Among them, in step S2, the low-frequency domain subband {AK} is diffused using the visual-linguistic cue and the diffusion model to achieve low-frequency domain enhancement of the low-light image, including driving appearance restoration based on the visual-linguistic cue {Tp, Tn}, where Tp represents a positive semantic cue and Tn represents a negative semantic cue; the diffusion model adjusts the generated result based on the visual-linguistic cue.

[0007] The diffusion model adjusts and generates results based on the visual-linguistic prompts, including the following steps: Feed the visual-linguistic cue {Tp, Tn} into the text encoder of the CLIP model to obtain the corresponding text feature vector {Fp, Fn}; The low-light image is input into the image encoder of the CLIP model to obtain the image feature vector Fi; The degree of match between the image and the prompt is evaluated based on comparing the distance between the text feature vector {Fp, Fn} and the image feature vector Fi.

[0008] In step S1, the low-light image and the normal-light image are decomposed by discrete wavelet transform based on multiple discrete wavelets to obtain four subbands {AK1, VK1, HK1, DK1} corresponding to the low-light image and four subbands {AK2, VK2, HK2, DK2} corresponding to the normal-light image; the high-frequency perception module processes the high-frequency domain subbands {VK, HK, DK} based on Fourier transform, including the following steps: The high-frequency perception module enhances the high-frequency sub-bands {VK1, HK1, DK1} of the low-light image; discrete Fourier transform is performed on the enhanced high-frequency sub-bands to achieve spatial-frequency domain conversion, and the detailed information of the high-frequency sub-bands is converted into a spectrum; The high-frequency perception module performs discrete Fourier transform on the high-frequency sub-bands {VK2, HK2, DK2} of the normal illumination image to obtain the frequency spectrum of the normal illumination image; A spectrum loss is obtained based on the enhanced high-frequency sub-band spectrum and the spectrum of the normal illumination image; and the enhanced high-frequency sub-band spectrum is optimized based on the spectrum loss.

[0009] The low-light image enhancement method can construct a loss function, which includes diffusion loss, multi-scale semantic guidance loss, spectrum loss and content reconstruction loss; the specific formula is: in, is the total loss value, is the diffusion loss, is the multi-scale semantic guided loss, is the spectrum loss, Reconstruct loss for content.

[0010] The diffusion loss is determined by a low-frequency subband {AK1} of the low-light image and a low-frequency subband {AK2} of the normal-light image, and the diffusion loss is minimized by minimizing the L2 distance between {AK1} and {AK2}; the L2 distance is the Euclidean distance between {AK1} and {AK2}.

[0011] Among them, the multi-scale semantic guidance loss is determined by the distance between the text feature vector {Fp, Fn} and the image feature vector Fi; the distance between {Fp, Fn} and Fi represents the level of semantic alignment; repeatedly calculating the distance between {Fp, Fn} and Fi based on different scales improves the alignment level of the generated image with the visual-linguistic cue {Tp, Tn}.

[0012] The content reconstruction loss includes MSE loss and SSIM loss; the MSE loss is the mean square error at the pixel level between the final enhanced image and the normal illumination image; the SSIM loss measures the structural similarity between the final enhanced image and the normal illumination image.

[0013] The second technical solution adopted in this application is: providing an electronic device, which includes: a memory and a processor coupled to each other, and the processor is used to execute program instructions stored in the memory to implement the low-light image enhancement method as described in any of the above items.

[0014] The third technical solution adopted in the present application is: providing a computer-readable storage medium, wherein the computer-readable storage medium stores program data, and the program data can be executed by a processor to implement the low-light image enhancement method as described in any one of the above items.

[0015] Due to the adoption of the above technical solution, the present application has at least one of the following beneficial effects compared with the prior art:

[0016] 1. By decomposing the image into the low-frequency domain {AK} and the high-frequency domain {VK, HK, DK}, the spatial dimension is greatly reduced, thereby reducing the computational complexity and resource consumption.

[0017] 2. The combination of visual-linguistic cues {T p, T n} and the diffusion model can effectively guide the enhancement process of the low-frequency subband {AK}, making the generated image more semantically consistent.

[0018] 3. The High-Frequency Perception Module (HFPM) performs Fourier transform on the high-frequency subbands {VK, HK, DK} to explore the high-frequency information of the degraded image from a frequency domain perspective and effectively restore the details and texture of the image.

[0019] 4. The loss function is divided into four parts: diffusion loss, multi-scale semantic guidance loss, spectrum loss, and content reconstruction loss. They are optimized from the dimensions of low-frequency domain, semantic alignment, high-frequency domain, and global content to ensure the comprehensiveness and high quality of the enhancement results. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without inventive efforts. Among them: Figure 1 A schematic diagram of a flow chart of an embodiment of a low-light image enhancement method provided by this application; Figure 2 This is a schematic structural diagram of an embodiment of an electronic device of the present application; Figure 3 This is a structural diagram of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It will be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some, rather than all, structures related to the present application are shown in the drawings. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0022] The terms "first," "second," and the like in this application are used to distinguish between different objects, not to describe a specific order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements and may include steps or elements not listed, or may include other steps or elements inherent to the process, method, product, or apparatus.

[0023] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0024] Low-light image enhancement technology has made significant progress, but unstable image quality restoration and unsatisfactory visual perception remain major challenges. To address the above problems, this application proposes a novel and robust low-light image enhancement method that utilizes CLIP-Fourier guided wavelet diffusion (CFWD); Figure 1 As shown, Figure 1 The following is a flowchart of an embodiment of a low-light image enhancement method provided in this application, including the following steps: Step S1: Based on the multiple discrete wavelet transform (K-DWT), the low-light image and the normal-light image are decomposed into subbands: {AK, VK, HK, DK}; where AK represents the low-frequency domain, VK, HK, and DK represent the high-frequency domain in the vertical, horizontal, and diagonal directions, respectively. The information of each subband is described in detail below: AK: Low-frequency domain, mainly contains the overall structure and brightness information of the image.

[0025] VK: High-frequency domain in the vertical direction, capturing the details and edge features of the image in the vertical direction.

[0026] HK: High-frequency domain in the horizontal direction, capturing the details and edge features of the image in the horizontal direction.

[0027] DK: High-frequency domain in the diagonal direction, capturing the details and edge features of the image in the diagonal direction.

[0028] The low-frequency domain contains the overall structure and brightness information of the image, while the high-frequency domain captures the details and edge features of the image; low-frequency domain enhancement can improve the overall brightness and contrast of the image; high-frequency domain enhancement can enhance the details and edge clarity of the image and avoid blurring or distortion; by processing the low-frequency domain and high-frequency domain separately, the relationship between overall structure and detail recovery can be better balanced.

[0029] Step S2: Build a multi-scale visual-linguistic guidance network, combining frequency-domain diffusion and multimodal methods for appearance guidance. Using visual-linguistic cues and a diffusion model, diffuse the low-frequency subband {AK} to achieve low-frequency enhancement of low-light images. The diffusion model focuses on enhancing the low-frequency subband {AK}, significantly improving the overall brightness and contrast of the image while ensuring semantic consistency and visual perceptual quality of the enhanced results. This method demonstrates excellent stability, generalization, and enhancement effectiveness.

[0030] Step S3: The high-frequency perception module processes the high-frequency subbands {VK, HK, DK} based on Fourier transform to achieve high-frequency domain enhancement of low-light images. The high-frequency perception module focuses on processing the high-frequency subbands {VK, HK, DK}, which can effectively restore the details and texture information in the image, making the enhanced image clearer and ensuring the naturalness and high quality of the enhancement result. Step S4: The enhanced low-frequency subband and high-frequency subband are subjected to inverse wavelet transform (IDWT) to obtain the final enhanced image. The function of the inverse wavelet transform is to fuse the low-frequency information and the high-frequency information together to restore the spatial structure of the image. By fusing the low-frequency and high-frequency information, the generated image appears more natural in the global scope, avoiding the incoordination problem caused by processing the low-frequency or high-frequency separately. The inverse wavelet transform can achieve a smooth transition between the low-frequency and high-frequency information, making the enhanced image more visually natural.

[0031] In summary, the low-light image enhancement method of this embodiment includes the following steps: S1: Based on multiple discrete wavelets, discrete wavelet transform decomposition is performed on the low-light image and the normal-light image to obtain subbands: {AK, VK, HK, DK}; S2: A multi-scale visual-linguistic guidance network is constructed, and appearance guidance is combined with frequency-domain diffusion and multimodal methods. The low-frequency subband {AK} is diffused using visual-linguistic cues and a diffusion model to achieve low-frequency domain enhancement of the low-light image; S3: A high-frequency perception module processes the high-frequency subband {VK, HK, DK} based on Fourier transform to achieve high-frequency domain enhancement of the low-light image; S4: The enhanced low-frequency subband and high-frequency subband are subjected to inverse wavelet transform to obtain the final enhanced image. By decomposing the image into the low-frequency domain {AK} and the high-frequency domain {VK, HK, DK}, the spatial dimension is significantly reduced, thereby reducing computational complexity and resource consumption.

[0032] In one embodiment, in step S2, the low-frequency domain subband {AK} is diffused using the visual-linguistic cue and the diffusion model to achieve low-frequency domain enhancement of the low-light image, including driving appearance restoration based on the visual-linguistic cue {Tp, Tn}, where Tp represents a positive semantic cue (e.g., bright, clear), and Tn represents a negative semantic cue (e.g., blur, noise). The visual-linguistic cue can be obtained in the following manner: Users directly provide positive and negative semantic hints through natural language.

[0033] Utilize machine learning algorithms to extract common visual-linguistic cues from historical data.

[0034] Combining image information with language information to generate prompts through multimodal models such as CLIP.

[0035] The diffusion model generates results based on visual-linguistic cues and includes the following steps: The visual-linguistic cue {Tp, Tn} is input into the text encoder of the CLIP model to obtain the corresponding text feature vector {Fp, Fn}; Fp represents the semantic feature of the positive cue, and Fn represents the semantic feature of the negative cue.

[0036] The low-light image is input into the image encoder of the CLIP model to obtain the image feature vector Fi; this vector represents the visual content of the low-light image.

[0037] The degree of match between the image and the prompt is evaluated by comparing the distance between the text feature vector {Fp, Fn} and the image feature vector Fi. The cosine similarity or Euclidean distance can be used to evaluate the degree of match between the image and the prompt. If the distance between Fi and Fp is small, it means that the current image is more in line with the requirements of the positive semantic prompt. If the distance between Fi and Fn is large, it means that the current image avoids the undesirable features described by the negative semantic prompt.

[0038] Based on the semantic matching results, the diffusion model dynamically adjusts the generation process; if the matching degree is low, the diffusion model will further optimize the enhancement results to better meet the requirements of semantic cues.

[0039] By comparing the distance between the text feature vector {Fp, Fn} and the image feature vector Fi, the diffusion model can accurately understand the semantic cues provided by the user and generate images that meet expectations; by freezing the latent space of the pre-trained CLIP model, the number of model parameters is reduced, and the stability and sampling quality of the diffusion model are improved.

[0040] In one embodiment, in step S1, discrete wavelet transform decomposition is performed on the low-light image and the normal-light image based on multiple discrete wavelets to obtain four subbands {AK1, VK1, HK1, DK1} corresponding to the low-light image and four subbands {AK2, VK2, HK2, DK2} corresponding to the normal-light image; and the high-frequency perception module processes the high-frequency subbands {VK, HK, DK} based on Fourier transform, including the following steps: The high-frequency perception module enhances the high-frequency subbands {VK1, HK1, DK1} of low-light images; it enhances the high-frequency subbands of low-light images through specific algorithms (such as nonlinear enhancement, filtering, etc.) to restore degraded details and texture information.

[0041] The enhanced high-frequency sub-bands are subjected to discrete Fourier transform to realize the conversion from spatial domain to frequency domain, and the detailed information of the high-frequency sub-bands is converted into spectrum; The high-frequency perception module performs discrete Fourier transform on the high-frequency sub-bands {VK2, HK2, DK2} of the normal illumination image to obtain the spectrum of the normal illumination image; A spectrum loss is obtained based on the enhanced high-frequency subband spectrum and the spectrum of the normal illumination image; the spectrum loss measures the difference between the enhanced high-frequency subband spectrum and the normal illumination image spectrum; and the enhanced high-frequency subband spectrum is optimized based on the spectrum loss.

[0042] The high-frequency perception module constructs a mixed frequency domain space, enhancing sensitivity to image structure and ensuring that the enhancement result is closer to the real scene; by comparing the difference between the enhanced high-frequency sub-band spectrum and the normal illumination image spectrum, the generated image is ensured to be highly consistent with the normal illumination image in terms of high-frequency information; the high-frequency perception module explores the high-frequency information of the degraded image from a frequency domain perspective. This method has strong generalization capabilities for different types of low-light images.

[0043] In one embodiment, the low-light image enhancement method may construct a loss function, which includes diffusion loss, multi-scale semantic guidance loss, spectrum loss, and content reconstruction loss. The specific formula is: in, is the total loss value, is the diffusion loss, is the multi-scale semantic guided loss, is the spectrum loss, Reconstruct loss for content.

[0044] The diffusion loss is determined by the low-frequency subband {AK1} of the low-light image and the low-frequency subband {AK2} of the normal-light image. The diffusion loss is minimized by minimizing the L2 distance between {AK1} and {AK2}. The L2 distance is the Euclidean distance between {AK1} and {AK2}. During the training process, the model parameters are adjusted to minimize the diffusion loss through optimization algorithms (such as gradient descent). , so that the low-frequency subband of the generated image is closer to the low-frequency subband of the normal illumination image.

[0045] Low-frequency subbands are closely related to the overall brightness of an image. Optimizing diffusion loss effectively improves the brightness consistency of the generated image, avoiding overly bright or dark conditions. Minimizing diffusion loss helps generate more natural and realistic enhanced images, bringing the visual effect closer to images under normal lighting conditions.

[0046] The multi-scale semantic guidance loss is determined by the distance between the text feature vector {Fp, Fn} and the image feature vector Fi; the distance between {Fp, Fn} and Fi represents the level of semantic alignment; repeatedly calculating the distance between {Fp, Fn} and Fi at different scales improves the alignment level between the generated image and the visual-linguistic cue {Tp, Tn}. During the training process, the model parameters are adjusted through the optimization algorithm to minimize the multi-scale semantic guidance loss. , so that the feature vector Fi of the generated image is closer to the positive prompt Fp and away from the negative prompt Fn.

[0047] The multi-scale semantically guided loss ensures that the generated image is more semantically consistent with the prompts provided by the user by comparing the distance between the text feature vector {Fp, Fn} and the image feature vector Fi; by optimizing the generation process through semantic guidance, the enhanced image is visually more natural and in line with human perception needs.

[0048] The content reconstruction loss includes MSE loss and SSIM loss; MSE loss is the mean square error at the pixel level between the final enhanced image and the normal illumination image; SSIM loss measures the structural similarity between the final enhanced image and the normal illumination image.

[0049] The MSE loss directly measures the pixel-level differences and can effectively reduce the brightness and color deviations between the enhanced image and the reference image. The SSIM loss focuses on the structural similarity of the image and can ensure that the enhanced image is consistent with the normal lighting image in terms of edges, textures, and other structural features.

[0050] The weighted sum of the MSE loss and the SSIM loss gives the final content reconstruction loss: is the mean square error, is the structural similarity error, α and β are weight coefficients used to balance the contributions of the two losses; it should be noted that in different embodiments, the values ​​of α and β may be different and are not limited here.

[0051] The combination of MSE loss and SSIM loss can simultaneously optimize the details and overall quality of the image, making the generated image more visually natural and in line with human perception needs; for example, the accuracy of brightness and color is ensured based on MSE loss, and the clarity of edges and textures is ensured based on SSIM loss.

[0052] For the above embodiment, this application provides an electronic device, see Figure 2 , Figure 2 This is a schematic diagram of the structure of an embodiment of an electronic device of the present application. The electronic device includes a memory and a processor, wherein the memory and the processor are coupled to each other, the memory stores program data, and the processor is configured to execute the program data to implement the steps of any embodiment of the low-light image enhancement method described above.

[0053] In this embodiment, the processor may also be referred to as a CPU (Central Processing Unit). The processor may be an integrated circuit chip having signal processing capabilities. The processor may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The general-purpose processor may be a microprocessor or any conventional processor.

[0054] The method of the above embodiment can be implemented in the form of a computer program, so this application proposes a computer readable storage medium, please refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of an embodiment of a computer-readable storage medium of the present application. The computer-readable storage medium stores program data that can be executed by a processor to implement the steps of any embodiment of the low-light image enhancement method described above.

[0055] The computer-readable storage medium in this embodiment can be a medium that can store program data, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or it can also be a server that stores the program data. The server can send the stored program data to other devices for execution, or it can also execute the stored program data itself.

[0056] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical functional division. In actual implementation, other division methods may be used, such as combining or integrating multiple units or components into another system, or ignoring or not implementing certain features.

[0057] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of this embodiment.

[0058] In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The above-mentioned integrated units may be implemented in the form of hardware or software functional units.

[0059] The above description is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A low-light image enhancement method, characterized in that: The steps include: S1: Based on multiple discrete wavelets, the low-light image and the normal-light image are decomposed by discrete wavelet transform to obtain subbands: {AK, VK, HK, DK}; where AK represents the low-frequency domain, VK, HK and DK represent the high-frequency domain in the vertical, horizontal and diagonal directions respectively; S2: Build a multi-scale visual-language guidance network, combine frequency domain diffusion and multimodal methods for appearance guidance, and use visual-language cues and diffusion models to diffuse the low-frequency subband {AK} to achieve low-frequency domain enhancement of low-light images; S3: The high-frequency perception module processes the high-frequency subband {VK, HK, DK} based on Fourier transform to achieve high-frequency domain enhancement of low-light images; S4: The enhanced low-frequency sub-band and high-frequency sub-band are subjected to inverse wavelet transformation to obtain a final enhanced image.

2. The low-light image enhancement method according to claim 1, characterized in that: In step S2, the low-frequency subband {AK} is diffused using the visual-linguistic cue and the diffusion model to achieve low-frequency domain enhancement of the low-light image, including driving appearance restoration based on the visual-linguistic cue {Tp, Tn}, Tp represents a positive semantic cue, and Tn represents a negative semantic cue; the diffusion model adjusts the generated result based on the visual-linguistic cue.

3. The low-light image enhancement method according to claim 2, characterized in that: The diffusion model generates results based on the visual-linguistic cue adjustment, including the following steps: Input the visual-linguistic cue {Tp, Tn} into the text encoder of the CLIP model to obtain the corresponding text feature vector {Fp, Fn}; The low-light image is input into the image encoder of the CLIP model to obtain the image feature vector Fi; The matching degree between the image and the hint is evaluated based on comparing the distance between the text feature vector {Fp, Fn} and the image feature vector Fi.

4. The low-light image enhancement method according to claim 1, characterized in that: In step S1, based on multiple discrete wavelets, the low-light image and the normal-light image are respectively decomposed by discrete wavelet transform to obtain four subbands {AK1, VK1, HK1, DK1} corresponding to the low-light image and four subbands {AK2, VK2, HK2, DK2} corresponding to the normal-light image; the high-frequency perception module processes the high-frequency domain subbands {VK, HK, DK} based on Fourier transform, including the following steps: The high-frequency perception module enhances the high-frequency sub-bands {VK1, HK1, DK1} of the low-light image; discrete Fourier transform is performed on the enhanced high-frequency sub-bands to achieve spatial domain-frequency domain conversion, and the detail information of the high-frequency sub-bands is converted into a spectrum; The high-frequency perception module performs discrete Fourier transform on the high-frequency sub-bands {VK2, HK2, DK2} of the normal illumination image to obtain the frequency spectrum of the normal illumination image; A spectrum loss is obtained based on the enhanced high-frequency sub-band spectrum and the spectrum of the normal illumination image; and the enhanced high-frequency sub-band spectrum is optimized based on the spectrum loss.

5. The low-light image enhancement method according to claim 4, characterized in that: The low-light image enhancement method can construct a loss function, which includes diffusion loss, multi-scale semantic guidance loss, spectrum loss and content reconstruction loss; the specific formula is: in, is the total loss value, is the diffusion loss, is the multi-scale semantic guided loss, is the spectrum loss, Reconstruction loss for content.

6. The low-light image enhancement method according to claim 5, characterized in that: The diffusion loss is determined by the low frequency subband {AK1} of the low light image and the low frequency subband {AK2} of the normal light image, and the diffusion loss is minimized based on minimizing the L2 distance between {AK1} and {AK2}; The L2 distance is the Euclidean distance between {AK1} and {AK2}.

7. The low-light image enhancement method according to claim 5, characterized in that: The multi-scale semantic guided loss is determined by the distance between the text feature vector {Fp, Fn} and the image feature vector Fi; the distance between {Fp, Fn} and Fi represents the level of semantic alignment; repeatedly calculating the distance between {Fp, Fn} and Fi at different scales improves the alignment level between the generated image and the visual-linguistic cue {Tp, Tn}.

8. The low-light image enhancement method according to claim 5, characterized in that: The content reconstruction loss includes MSE loss and SSIM loss; the MSE loss is the mean square error at the pixel level between the final enhanced image and the normal illumination image; the SSIM loss measures the structural similarity between the final enhanced image and the normal illumination image.

9. An electronic device, characterized in that: The electronic device comprises: a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the low-light image enhancement method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program data, and the program data can be executed by a processor to implement the low-light image enhancement method according to any one of claims 1 to 8.

Images