Language feature fused image color migration network weak light image enhancement method

Abstract

An image color migration network weak light image enhancement method fusing language features comprises the steps that S1, based on the rectification theory, a low light content image and a reference style image are modeled into RGB space three-dimensional probability distribution, and smooth migration of color distribution is achieved through an ordinary differential equation and linear interpolation; s2, decomposing the low-light image based on the Retinex theory, and respectively extracting brightness details and color information by a double-branch encoder to generate a super-resolution image; s3, analyzing a user text instruction, decomposing a region descriptor and an intensity coefficient, generating an enhanced mask, and realizing local region optimization; and S4, pre-training related modules in stages, and performing end-to-end fine tuning through a composite loss function to enable color distribution, structural details and style reference of an output image to be consistent. Compared with the traditional method, the method provided by the invention can realize personalized enhancement and improve the weak light image quality and the style migration effect, and has the remarkable advantages of flexible input, high fidelity, no artifact, detail reservation, strong noise suppression capability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to an image color transfer network low-light image enhancement method fusing language features. BACKGROUND

[0002] Low-light image enhancement (LLIE) plays a crucial role in realizing downstream computer vision tasks and has attracted extensive attention in the academic community. However, LLIE is still a highly ill-posed problem because there is a complex and ambiguous mapping relationship between low-light images and well-exposed images. At the same time, traditional color style transfer techniques commonly used for photo editing are prone to artifacts, low image fidelity, and limited interpretability.

[0003] Low-light image enhancement (LLIE) aims to improve the visibility, contrast, and perceptual quality of images captured under dim lighting conditions while preserving natural details and color fidelity. LLIE is a key preprocessing step for various downstream computer vision applications such as image recognition, object detection, traffic monitoring, semantic segmentation, etc. Therefore, LLIE has become a cross-disciplinary research hotspot and has attracted extensive attention.

[0004] Traditional distribution mapping-based LLIE methods include histogram equalization (HE) and s-curve-based methods. However, histogram equalization (HE)-based methods are only effective in images with uniform lighting, while s-curve-based methods often result in color distortion. Model-based optimization methods based on the Retinex theory decompose images into reflectance and illumination components. Although these methods use regularization terms to guide the decomposition, they are prone to artifacts, blurred details, and unsaturated colors due to reliance on pre-determined design priors. In recent years, with the advent of deep learning, end-to-end LLIE methods have achieved remarkable success in image restoration and denoising (such as FEUNet, DRANet, PhotoWCT, RetinexNet, KinD++, etc.). To improve representation capacity, a series of ar architectures have been proposed that utilize deep convolutional neural networks (cnn), warping, Mamba, generative adversarial networks (gan), and diffusion models. In addition, LLIE still has inherent ill-posedness because the mapping from low-light to well-exposed images is one-to-many. Many existing methods over-rely on pixel-level reconstruction loss and fail to capture the complex conditional distribution of well-exposed images, leading to overexposure, color abnormalities, and structural artifacts.

[0005] Both supervised and unsupervised paradigms attempt to incorporate structural priors to improve enhancement quality. In unsupervised learning, models like CUT leverage contrastive learning operations on raw image patches and achieve one-sided translation in unpaired image-to-image settings. DRBN proposes a semi-supervised approach to LLIE, benefiting from both paired and unpaired data. Recently, Yang et al. proposed DKP, an unsupervised kernel estimation model for super-resolution based on Markov Chain Monte Carlo sampling. However, these methods still rely on artificially designed loss functions and lack consistency with human perception priors, resulting in limited interpretability and suboptimal visual fidelity.

[0006] Color style transfer, which matches a reference style by altering the emotional and perceptual tonalities of an image, is widely used in photo editing applications to correct lighting, unify the aesthetics of an image set, and emphasize key content. Commercial solutions often rely on image filters or lookup tables (LUTs). For example, Gatys et al. proposed a pioneering method based on iterative optimization of content and style reconstruction losses. Later, feedforward networks were introduced to further reduce computational costs. However, their encoder-decoder architecture often introduces artifacts and compromises overall image quality. Additionally, these methods lack human interpretability and often neglect image denoising or super-resolution tasks, which are crucial for obtaining high-fidelity results in low-light image enhancement scenarios.

[0007] Overall, existing methods have the following problems: 1. Ill-posedness of traditional LLIE: Existing LLIE methods have a complex and ambiguous mapping relationship between low-light images and well-exposed images, making it a highly ill-posed problem. This causes many existing methods to rely excessively on pixel-level reconstruction losses, failing to capture the complex conditional distribution of well-exposed images, which can easily lead to improper exposure, color abnormalities, and structural artifacts.

[0008] 2. Inherent limitations of traditional color style transfer techniques: Conventional color style transfer techniques, although commonly used in photo editing, are prone to artifacts, low image fidelity, and limited interpretability. Their encoder-decoder architecture often introduces artifacts and compromises overall image quality. Additionally, these methods lack human interpretability and often neglect image denoising or super-resolution tasks, which are crucial for obtaining high-fidelity results in LLIE scenarios.

[0009] 3. Lack of personalization and semantic perception: Existing LLIE methods and color style transfer techniques are usually global, making it difficult to achieve fine control over local regions of an image and adjust according to users' specific intentions. This limits users' ability to customize enhancements according to their personal preferences and aesthetic goals.

[0010] 4. Color extraction process is susceptible to noise: In low light conditions, the soft color features inherent to the image and the fine details that are obscured due to insufficient illumination make the color extraction process extremely susceptible to the introduction of noise, which can adversely affect the final style transfer performance.

[0011] It should be noted that the information disclosed in the above background section is only for understanding the background of the present application, and therefore can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0012] The main purpose of the present application is to overcome the defects existing in the background art, and provide a weak light image enhancement method of image color transfer network fusing language features.

[0013] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: A weak light image enhancement method of image color transfer network fusing language features, comprising the following steps: S1, color distribution modeling and transmission: based on the rectification theory, the input low light content image and the reference style image are modeled as three-dimensional probability distribution in RGB space, the transmission process between the distributions is described by ordinary differential equation, the intermediate state is constructed by linear interpolation and the transmission mapping function is fitted, realizing the smooth migration of color distribution from content image to style image; S2, super-resolution color feature extraction: based on the Retinex theory, the low light image is decomposed into illumination component and reflectivity component, and a double-branch encoder is used to extract brightness details and color information respectively; wherein the reflectivity branch integrates U-Net structure and bilateral filter for noise suppression and edge preservation, and the illumination branch adopts residual convolution network to enhance structural details, and generates a super-resolution image through cross-branch feature fusion; S3, text-driven regional enhancement: analyze the user text instruction, and decompose the region descriptor and intensity coefficient through the language model; generate a spatially accurate enhancement mask combining the reflectivity component and the region descriptor, control the rectification transmission process based on the mask and the intensity coefficient, realize the iterative optimization and adaptive mixing of local region; S4, modular joint training: pre-train the Retinex decomposition module, super-resolution module and rectification encoder in stages, and fine-tune end-to-end through a composite loss function, so that the color distribution, structural details and style reference of the output image are consistent.

[0014] Further, step S1 specifically includes the following process: The pixel samples of the content image and the style image are respectively regarded as corresponding random variables; The intermediate state variable between the two is constructed by linear interpolation in the time dimension; A system of ordinary differential equations is defined to describe the distribution transport process, and the transport function is optimized by minimizing the fitting loss between the transport function and the intermediate state variable; A latent distribution is introduced to bridge the content image distribution and the style image distribution, and the color distribution transport from the content image to the style image is achieved by combining the forward sampling of the content image distribution and the backward sampling of the style image distribution.

[0015] Further, the transport mapping and its inverse mapping are realized by integrating the transport function: The forward transport mapping is obtained by integrating the transport function along the time domain in the positive direction, and is used for transporting from the content image distribution to the intermediate latent distribution; The backward transport mapping is obtained by integrating the transport function along the time domain in the reverse direction, and is used for transporting from the intermediate latent distribution to the style image distribution.

[0016] Further, in step S2, the specific operation process of the double-branch encoder includes: Reflectance branch: U-Net structure is used for downsampling and upsampling operations, and an edge-preserving denoising module is inserted between each downsampling layer and the corresponding upsampling layer to suppress the noise introduced in the downsampling process; Illuminance branch: a residual network group composed of multiple layers of convolution and activation function is used to extract high-frequency brightness detail features from the illuminance image; In the first layer and the last layer of the U-Net, an element-wise multiplication operation of the illuminance branch output and the reflectance branch output is introduced, so that the reflectance branch can adaptively focus on the chroma information; The outputs of the double-branch are fused by element-wise multiplication, and then weighted summed with the input image to generate a super-resolution image after tail processing.

[0017] Further, the edge-preserving denoising module uses a bilateral filter to bridge each downsample-upsample pair of the U-Net, suppressing the noise introduced in the downsampling process while maintaining the sharpness of the image edges.

[0018] Further, in step S2, the super-resolution image is mapped to the weight parameters of the rectified transport function by a convolutional neural network, establishing the association between the super-resolution image features and the rectified transport process.

[0019] Further, step S3 specifically includes the following processing process: The language model is used to analyze the user input text instruction, and outputs the intensity parameter for regulating the enhancement degree and the semantic region description for positioning the target region; The segmentation model combines the reflectance component obtained from the Retinex encoder to generate a spatially accurate binary mask according to a semantic region description, so as to identify a local region that needs to be enhanced; For the region covered by the mask, a rectified transmission process under the control of the intensity parameter is applied to realize the style transfer and enhancement of the region. By iteratively optimizing the rectified transmission parameters, the enhancement effect at the boundary of the mask is smoothly transitioned, and obvious region segmentation traces are avoided.

[0020] Further, the step S4 of the staged pre-training specifically includes the following stages in sequence: The first stage: taking the decomposition loss of reflectance and illumination components as the target, the dual-branch encoder is trained so that the reflectance and illumination components obtained by decomposition can accurately reconstruct the weak light image input; The second stage: the training image is down-sampled and then up-sampled to the original size to simulate a blurred input, and the super-resolution module is trained with a multi-constrained reconstruction loss as the target, so that the module can reconstruct a high-resolution image from the blurred input; The third stage: the pre-trained super-resolution module is combined with the rectified encoder, and an end-to-end fine-tuning is performed with a composite loss function integrating multiple losses as the target to optimize the overall network parameters.

[0021] Further, the composite loss function specifically includes the following components: Encoding loss: by constraining the difference between the predicted value and the true value of the transmission function, the fitting accuracy of the transmission function is ensured; Style loss: by matching the deep features of the generated image and the reference style image, the consistency of the two in style is constrained; Content loss: by constraining the pixel-level difference between the generated image and the content image, the structural content consistency of the image is maintained; Reconstruction loss: by integrating the constraints on image gradient details, the regularization constraints on color space and the measurement constraints on structural similarity, the details and quality of the generated image are ensured.

[0022] A computer program product comprising a computer program which, when executed by a processor, implements the image color transfer network weak light image enhancement method of fusing language features.

[0023] The present application has the following advantages: The present application is directed to the mapping relationship between the complex blur of low-light images and well-exposed images in low-light image enhancement (LLIE), and the problems of traditional color style transfer technology, such as easy to produce artifacts, low image fidelity, and limited explainability. A low-light image enhancement method of image color transfer network fusing language features is proposed, and a text-driven color style transfer network (TCST-N) is constructed.

[0024] The method is based on text-driven and rectified flow theory, and guides local low-light enhancement and global color style transfer through natural language description. By regarding images as probability distribution in RGB space, using ordinary differential equations and linear interpolation to realize smooth transmission of color distribution, and combining LISA model (an image segmentation model) to realize accurate local area control, personalized and semantic-aware image enhancement is effectively achieved. At the same time, based on Retinex theory and deep learning, color details are extracted by using U-Net, bilateral denoising module and residual convolutional neural network, combined with VGG-based style loss function and modified flow coding loss, to realize more accurate and detailed color extraction and noise suppression, and improve the realism and consistency of style transfer. Through the construction of multi-exposure gradient (MEG) dataset, the superior performance and generalization ability in multiple challenging scenarios are verified by using modular training paradigm and composite loss function.

[0025] The important feature of the present application is that low-light image enhancement (LLIE) and color style transfer (CST) are deeply integrated rather than processed independently. The rectified flow theory is used to model the color distribution flow conversion from low-light images to well-exposed images, and a continuous mapping between images is established. In the encoder, the Retinex theory is used to decompose the low-light image into illumination and reflectance components, realizing independent color information extraction and brightness detail enhancement. The entire system is based on deep convolutional neural networks (CNNs), U-Net and other deep learning architectures for image feature extraction, conversion and generation.

[0026] Compared with existing methods, the present application has the following significant advantages: high input flexibility, supporting multiple inputs such as text instructions and reference style images, users can guide local low-light enhancement through natural language description; it can provide high-fidelity and artifact-free color style transfer, establish continuous mapping with the help of rectified flow theory, and ensure that the RGB distribution of the output image is closely consistent with the style image by combining related loss functions; it has excellent detail preservation and noise suppression capability, and its super-resolution module accurately extracts color details by combining multiple structures, and bilateral filter effectively reduces down-sampling noise and maintains edge sharpness.

[0027] In addition to low-light enhancement, the method of the present application has been widely verified in multiple challenging scenarios such as artifact coloring, artistic style transfer, infrared-to-visible spectrum conversion, and underwater image enhancement, and has shown good performance. It has a wide range of applications, can significantly enhance the visibility, contrast, and perceptual quality of surveillance images while preserving natural details and color fidelity in the surveillance and security fields; in the e-commerce field, it allows users to achieve local low-light enhancement and global lighting adjustment through natural language instructions, providing a highly customized experience; in the field of digitalization of cultural relics, it can greatly reduce the dependence on hardware equipment and manual operation for obtaining low-light images.

[0028] Other benefits of the embodiments of the present application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The architecture diagram of the text-driven color style transfer network (TCSTNet) of the embodiments of the present application.

[0030] Figure 2 The low-light enhancement result diagram of three content images and nine style images of the embodiments of the present application.

[0031] Figure 3 The low-light enhancement effect diagram of a single text command at 20%, 40%, 60%, and 80% intensity of the embodiments of the present application.

[0032] Figure 4 The color transfer result diagram compared with the style transfer method of the embodiments of the present application.

[0033] Figure 5 The low-light enhancement effect diagram compared with the style transfer method of the rain scene of the embodiments of the present application.

[0034] Figure 6 The low-light enhancement effect diagram compared with the style transfer method of the night scene of the embodiments of the present application.

[0035] Figure 7 The low-light enhancement result diagram compared with the style transfer method in the backlight scene of the embodiments of the present application.

[0036] Figure 8 The result diagram compared with the low-light enhancement method in the sunny scene of the embodiments of the present application.

[0037] Figure 9 The result diagram compared with the low-light enhancement method in the rainy scene of the embodiments of the present application.

[0038] Figure 10 The low-light enhancement method comparison result diagram in the cloudy scene of the embodiments of the present application.

[0039] Figure 11 Result image of the super-resolution module of the embodiment of the present application.

[0040] Figure 12 Result image of the ablation study of the embodiment of the present application.

[0041] Figure 13 The overall flow chart of the weak light image enhancement method of the image color transfer network fusing language features of the present application. DETAILED DESCRIPTION

[0042] The embodiments of the present application are described in detail below. It should be emphasized that the following description is merely exemplary, not intended to limit the scope of the present application and its applications.

[0043] The present application aims to solve the problems of complex and fuzzy mapping relationship in low light image enhancement and the artifacts and low fidelity of traditional color style transfer technology, and proposes a weak light image enhancement method of image color transfer network fusing language features, constructs a text-driven color style transfer network (referred to as TCST-N), which is based on rectified flow and Retinex theory, and realizes personalized enhancement combined with deep learning, has the advantages of flexible input, high fidelity without artifacts, strong ability of detail preservation and noise suppression, etc.

[0044] The embodiment of the present application provides a weak light image enhancement method of image color transfer network fusing language features (the overall flow chart is shown in Figure 13 ), which comprises the following steps: Step S1, color distribution modeling and transmission: based on the rectified theory, the input low light content image and the reference style image are modeled as three-dimensional probability distribution in RGB space, the transmission process between the distributions is described by ordinary differential equation, the intermediate state is constructed by linear interpolation and the transmission mapping function is fitted, realizing the smooth transfer of color distribution from content image to style image.

[0045] In some embodiments, step S1 specifically comprises the following process: the pixel samples of the content image and the style image are respectively regarded as corresponding random variables; the intermediate state variables between the two are constructed by linear interpolation in the time dimension; the ordinary differential equation describing the distribution transmission process is defined, and the transmission function is optimized by minimizing the fitting loss between the transmission function and the intermediate state variables; the latent distribution obeying uniform distribution on the uniform interval is introduced to bridge the content image distribution and the style image distribution, and the color distribution transmission from the content image to the style image is realized by the combination of forward sampling of the content image distribution and backward sampling of the style image distribution.

[0046] In some embodiments, the transport mapping and its inverse mapping are implemented by integration of the transport function: the forward transport mapping is obtained by forward integration of the transport function along the time domain, for transporting from the content image distribution to the intermediate latent distribution; the backward transport mapping is obtained by reverse integration of the transport function along the time domain, for transporting from the intermediate latent distribution to the style image distribution.

[0047] Step S2, super-resolution color feature extraction: based on the Retinex theory, the low-light image is decomposed into an illumination component and a reflectance component, and a double-branch encoder is used to extract brightness details and color information respectively; wherein the reflectance branch integrates a U-Net structure and a bilateral filter for noise suppression and edge preservation, and the illumination branch uses a residual convolutional network to enhance structural details, and generates a super-resolution image through cross-branch feature fusion.

[0048] In some embodiments, the specific operation process of the double-branch encoder in step S2 includes: reflectance branch: U-Net structure is used for downsampling and upsampling operations, and an edge-preserving denoising module is inserted between each downsampling layer and the corresponding upsampling layer to suppress the noise introduced in the downsampling process; illumination branch: a residual network group composed of multiple convolutional layers and activation functions is used to extract high-frequency brightness detail features from the illumination image; at the same time, element-wise multiplication operations of the output of the illumination branch and the output of the reflectance branch are introduced at the first layer and the last layer of the U-Net, so that the reflectance branch can adaptively focus on the chroma information; the outputs of the double-branch are fused through element-wise multiplication, and then weighted summation is performed with the input image, and a super-resolution image is generated after tail processing.

[0049] In some embodiments, the edge-preserving denoising module uses a bilateral filter to bridge each downsample-upsample pair of the U-Net, while suppressing the noise introduced in the downsampling process and preserving the sharpness of the image edges.

[0050] In some embodiments, the super-resolution image is mapped to the weight parameters of the rectifying transport function through a convolutional neural network, establishing the association between the super-resolution image features and the rectifying transport process.

[0051] Step S3, text-driven region enhancement: parse the user text instructions, and decompose the region descriptors and intensity coefficients through a language model; generate a spatially accurate enhancement mask combining the reflectance component and the region descriptors, and control the rectifying transport process based on the mask and the intensity coefficients to realize the iterative optimization and adaptive mixing of local regions.

[0052] In some embodiments, step S3 specifically comprises the following processing process: parsing the text instruction input by the user by using a language model, outputting an intensity parameter for regulating the enhancement degree and a semantic region description for positioning the target region; a segmentation model combines the reflectance component obtained from the Retinex encoder to generate a spatially accurate binary mask according to the semantic region description to identify the local region that needs to be enhanced; for the region covered by the mask, a rectified transmission process under the regulation of the intensity parameter is applied to realize the style transfer and enhancement of the region; and the rectified transmission parameter is iteratively optimized to make the enhancement effect at the mask boundary smoothly transition to avoid obvious region segmentation traces.

[0053] Step S4, modular joint training: pre-training the Retinex decomposition module, the super-resolution module and the rectified encoder in stages, and end-to-end fine-tuning through a composite loss function to make the color distribution, structural details and style of the output image consistent with the reference.

[0054] In some embodiments, the staged pre-training of step S4 specifically comprises the following stages in sequence: First stage: train the dual-branch encoder to make the decomposed reflectance and illumination components accurately reconstruct the input low-light image, with the decomposition loss of reflectance and illumination components as the target; Second stage: downsample the training image and then upscale it to the original size to simulate blurred input, train the super-resolution module to make it reconstruct a high-resolution image from the blurred input, with a multi-constrained reconstruction loss as the target; Third stage: jointly train the pre-trained super-resolution module and the rectified encoder to optimize the overall network parameters with an end-to-end fine-tuning of a composite loss function that integrates multiple losses as the target.

[0055] In some embodiments, the composite loss function specifically includes the following components: encoding loss: by constraining the difference between the predicted value and the true value of the transmission function, the fitting accuracy of the transmission function is ensured; style loss: by matching the deep features of the generated image and the reference style image, the consistency of the two in style is constrained; content loss: by constraining the pixel-level difference between the generated image and the content image, the structural content consistency of the image is maintained; reconstruction loss: by integrating the constraints on image gradient details, the regularization constraints on color space and the measurement constraints on structural similarity, the details and quality of the generated image are ensured.

[0056] The method of the present application solves the ill-posed mapping problem and the artifact defect of color style transfer in low-light image enhancement by fusing rectilinear flow theory, Retinex decomposition and text-driven mechanism. Specifically, first, the RGB space probability distribution modeling is adopted to describe the continuous transmission mapping between low-light images and normally exposed images through ordinary differential equations, which overcomes the exposure abnormalities and structural artifacts caused by fuzzy mapping in traditional methods; second, the double-branch encoder design based on Retinex theory is used to extract noise-robust reflectance features by using U-Net and bilateral filter, and the residual network is used to enhance the illumination details, which significantly improves the color fidelity and edge definition while suppressing noise; third, the text instruction driven local enhancement mechanism is created, which analyzes the semantic region description and intensity parameters through the language model, and combines the reflectance component to generate a spatially accurate mask, realizing user customizable regional adaptive optimization and breaking through the limitation of global processing lacking personalized control; fourth, the modular training paradigm and composite loss function are optimized, which integrates transmission coding loss, deep feature style matching, pixel-level content constraint and multi-scale reconstruction regularization, to ensure the high consistency of the output image in color distribution, structural details and style reference. Finally, the input flexibility (supporting natural language instructions), image quality extraction (high fidelity without artifacts), controllability (semantic level regional adjustment) and diversity (multi-scene generalization) are comprehensively improved, which is suitable for monitoring security, e-commerce, digital cultural relics and other fields that require high-precision low-light enhancement.

[0057] The specific embodiments of the present application, algorithm examples and experimental verification are further described below.

[0058] The present application proposes an image color transfer network low-light image enhancement method combined with language features. In order to overcome the limitations and deficiencies of existing methods, the method of the present application constructs a text-driven color style transfer network TCSTNet, which is a text-driven color style transfer network specially designed for LLIE. Figure 1 The text-driven color style transfer network TCSTNet is shown. TCSTNet realizes local low-light enhancement through natural language input and global illumination adjustment through reference-based color style transfer, embodying the vision of "natural language guided image adjustment". The method of the present application is evaluated on multiple benchmark datasets, including a newly constructed multi-exposure gradient (MEG) dataset (which is composed of aligned images of different exposures of the same scene), and compared with several state-of-the-art (SOTA) baseline methods. In addition, the ablation study further verifies the contribution of each proposed module. Based on the foregoing, the important innovative contributions of the present application include:

[0059] New task of text-driven low-light image enhancement: The invention introduces a personalized LLIE task that uses natural language descriptions to guide local low-light enhancement and combines it with global color style transfer to meet different user preferences and aesthetic goals.

[0060] Enhanced color style transfer model based on rectified flow: Based on the theory of rectified flow, the invention establishes a continuous mapping between the content image, latent image, style image and stylized image. On the basis of coding loss, the model combines style and content loss to ensure that the RGB distribution of the output image is closely consistent with the RGB distribution of the style image, thereby improving the realism and consistency.

[0061] Color extraction encoder based on super-resolution: The invention designs an encoder that combines U-Net with bilateral denoising modules and residual convolutional neural networks (CNN) to preserve image details. The final transformation parameters for the rectified flow are predicted using EfficientNet. This design can effectively suppress noise while achieving more accurate and detailed color extraction.

[0062] Low-light enhancement pipeline based on rectification Under the guidance of the concept of color style transfer, the low-light enhancement task aims to produce stylized results, allowing users to achieve customized enhancement effects according to the preferred visual style. The rectified flow is introduced to solve the color transfer problem and good results are obtained. According to the definition of rectified flow, the image is regarded as a continuous three-dimensional probability distribution in RGB space. Then, the random variable X 0 ~ π 0 And X 1 Pi1 refers to the pixel samples taken from the content image and the target image, so the definition of the style transfer problem can be given as

[0063]

[0064] In the formula, T is the transport mapping between the above two distributions, J T (x) is the mapping obtained at the sample point x T The Jacobian matrix of the mapping tThe flow on ∈ [0, 1] is the step size, which is a factor of numerical stability control in style transfer problems, so the size of the training gradient is used to balance the stability and convergence speed of the training process. The step size can be adjusted inversely proportional to the size of the gradient, ensuring that a smaller step size is used in the large gradient area and a larger step size is used in the small gradient area.

[0065] wherein is the norm of the current gradient, is a small constant to prevent division by zero errors, and a is a constant with a value of 10 -2 , which controls the adjustment amplitude.

[0066]

[0067] It completes the transportation from to Z1 as linearly as possible through the path in the direction of Z 0 Therefore, linear interpolation between X 0 and X 1 is introduced, and the transfer process is represented as

[0068]

[0069] wherein Xt is the intermediate state, t is the interpolation step. Then, the transfer v ( Z t , t) can be fitted by a neural network to minimize the following loss function

[0070]

[0071] The weights of the neural network v ψ ( Z t , t) are denoted as e ψ。 The encoder Enc (X ψ )= e ψ defines the function weights e ψ of the image X ψ , which will be discussed in the next subsection. The rectified flow has the non-crossing property, which means that the ODE: d Z t = v ( Z tThe solution of the equation is unique and the minimum cost can be achieved in transportation. From this, the transportation graph can be presented T

[0072]

[0073] In addition, the objective in equation (4) is time-symmetric, so the backward sampling can be calculated as is denoted as

[0074] Inspired by the time-symmetric property of rectification, a general latent distribution U is introduced to bridge the content image and style image. U The existence of can be given as: if the random variable X ~ π, X∈R has a monotonically increasing cumulative distribution function CDF

[0075]

[0076] then the transformed variable U = F(X) obeys a uniform distribution on [0,1]. Therefore, the transfer from the color distribution X 0 and X 1 to U is denoted as T 0 and T 1 respectively. Thus, the color transfer process can be denoted as

[0077]

[0078] where T 0 is the content image distribution X 0 Forward sampling of , is the style image distribution X 1 Backward sampling of and the stylized image distribution is X 01 。

[0079] Retinex-based super-resolution encoder According to the rectified pipeline, the encoder Enc (X ψThe color distribution of low-light images must be accurately extracted and rectified. Due to the soft color feature inherent in low-light conditions, and the fine details that are easily obscured by insufficient lighting, the color extraction process is susceptible to the introduction of noise, which can adversely affect the final style transfer performance.

[0080] Therefore, the low-light image is decomposed into illumination and reflectance components in conjunction with the Retinex theory. This separation allows independent extraction of color information (reflectance) and brightness details (illumination), thus facilitating more effective detail enhancement in subsequent processing stages. The theoretical basis of Retinex is the trichromatic theory and color constancy. This means that the color of an object is determined by its reflection characteristics of red, green, and blue light, rather than the absolute intensity of the reflected light. At the same time, the color of an object remains consistent and is not affected by uneven lighting. The Retinex model utilizes this color constancy to achieve a balance between dynamic range compression, edge enhancement, and color constancy. Therefore, it can adaptively enhance various types of images, unlike traditional methods that can only enhance specific features of an image.

[0081] However, the classic Retinex operation causes color distortion in the RGB space, which utilizes the maximum value of the RGB channels to obtain the initial illumination L 0 Then the reflectance is calculated by element-wise division R 0 .

[0082] Therefore, a convolutional neural network module composed of a full CNN is introduced to simultaneously learn the illumination and reflectance. The module consists of three Conv+LeakyReLU layers followed by a Conv+ReLU layer. To train the Retinex network, the loss function is defined as:

[0083]

[0084] where is the output image of the network, is the pixel value of the input image in color channel c, R 0 and L 0 are the reflectance component and the illumination component in the Retinex network, respectively, µ is a hyperparameter, c ∈{R,G,B} is the RGB channel, ||·||1 and ||·||F F denote l 1 the L1 norm and the F norm, respectively.

[0085] For Figure 1 ​The illumination branch in the middle, after passing through a head convolutional network, proposes a residual network group containing 8 Conv+ReLUs to extract detailed information from the illumination image. The reflection branch is composed of a head convolutional layer and a U-Net layer, which enhances the extraction of color information from low-light images. The input channel size of the down-sampling layer of the U-Net is 64, 128, 256, and 512, while the corresponding up-sampling layer is the opposite (512, 256, 128, 64). Bilateral filter bridges each down-sampling-up-sampling pair, which weakens the down-sampling noise before feature reconstruction, and its denoising performance is better than other methods such as Gaussian filter, achieving denoising while maintaining edge sharpness. In the first and last layers of the U-Net, element-wise multiplication with the illumination branch is introduced, allowing the reflection branch to adaptively focus on chrominance information under different lighting conditions.

[0086] The two branches are merged through element-wise multiplication, then weighted with the input image, and then a tail Conv+ReLU, generating the final super-resolution image with enhanced structural details. Then the Efficient-Net finally establishes the mapping relationship between the super-resolution image and the rectification parameter eψ.

[0087] Text-based regional image enhancement module Recognizing the subjectivity of low-light enhancement preferences and the need for localized style control, the framework of the present invention introduces a text command-based style transfer mechanism that supports specific region conversion. GPT parses the input text command and decomposes the style transfer parameters into: (i) an intensity coefficient for rectification adjustment, (ii) a target region descriptor for the LISA segmentation model (the LISA model is pre-trained based on the CLIP architecture, aligning text instructions with image region features through contrastive learning; the high-frequency edge information of the reflectance component is input into LISA as a spatial prior, constraining the mask generation to cover only connected regions with consistent chrominance). Notably, LISA also processes the reflection mapping from the Retinex encoder to generate spatially accurate masks.

[0088] For example, the user inputs "please enhance the left building a little bit", the "left building" segment will be processed in the LISA model and generate a mask, and the "a little bit" segment will be converted to 10% or 20% enhancement intensity by GPT.

[0089] If the user further instructs "make the right building a little brighter like the reference", the system can adjust the enhancement parameters through iterative rectification optimization and region-aware blending to finally generate a personalized low-light enhancement image that meets complex style requirements.

[0090] Training method and loss design Considering the inherent complexity of the model, a modular training paradigm is adopted, where each component is pretrained separately before the final end-to-end fine-tuning. This staged approach significantly speeds up the convergence compared to joint training from scratch. The Retinex module is first trained to obtain the reflectance and illumination images at the origin, whose loss function is shown in equation (9).

[0091] Then, the super-resolution module based on Retinex theory is trained. The training images are compressed to 25% of the resolution (4x compression), followed by upsampling to the original size, generating artificially blurred inputs. The super-resolution capability of the network is quantitatively verified by comparing the output with the corresponding high-resolution ground truth image using a combined index.

[0092] The loss function is:

[0093] where: The weights are determined by grid search to find the optimal value.

[0094] is l 1 The norm loss, which enforces pixel-level content consistency between the super-resolved image and the ground truth image.

[0095] is the gradient magnitude loss, which preserves high-frequency edge details by Sobel operator constraint.

[0096] is the color space loss, which uses LAB space regularization to prevent color distortion in the enhanced output.

[0097] is the structural similarity (SSIM) loss, which maintains global image statistics and contrast relationships.

[0098] λ L1 , λ grad , λ color , and λ SSIM are the weights of each loss term, respectively.

[0099] The expression of the loss function is:

[0100] where C, H, and W are the channel size, height size, and width size, respectively, I output ( c , h , w ) represent the output image in the channel cheight h width w pixel value at height I target ( c , h , w ) represent the pixel values of the target image at channel c height h width w

[0101]

[0102] ▽ x I output and ▽ y I output represent the gradients in the directions of I output x and y , ▽ x I target and ▽ y I target represent the gradients in the directions of I target x and y , ▽ x and ▽ y represent the gradient operators in the directions of x and y , L out , a out , b out represent the L, a, b channel values of the output image in the Lab color space, L tgt , a tgt , b tgt represent the L, a, b channel values of the target image in the Lab color space, L represent the lightness of the image, R , G , B represent the values of the corresponding color channels in the RGB color space, L SSIM is a loss function based on structural similarity SSIM,

[0103] where , µ is the mean value, and σ​​​x , σ y is the standard deviation, σ xy is the covariance, C 1 and C 2 is a stabilization constant, whose optimal value is determined by experiments to achieve a balance between noise suppression and edge preservation.

[0104] The pre-trained Retinex-based super-resolution network is combined with EfficientNet to fine-tune the final encoder parameters. This stage has a self-adaptive weighted compound loss function formula as follows:

[0105]

[0106] where is the RetiConv loss, is the rectified encoding loss, is the VGG-style loss, is l 1 normative content loss, λ RetiConv , λ enc , λ style and λ content are the weights of each loss term. The specific expression is as follows:

[0107]

[0108] where v ψ ( X t , t) is the ODE vector field predicted by the encoder, v true is true value.

[0109]

[0110] where φ VGG represents the features from the fourth layer of VGG19, I style is the reference style image, I output is the final output image. This term enforces high-level style transfer through deep feature matching.

[0111] The ground truth v true of each image, a two-layer MLP with 1024 hidden neurons and tanh activation function is trained to generate the image parameter pair of each sample.

[0112] Experiments Datasets and metrics Three datasets were trained and validated: two widely used low-light datasets (LOL and LOL-V2 datasets) and a multi-exposure gradient (MEG) dataset collected by the inventors. The LOL dataset contains 500 pairs of images, while the LOL-V2 dataset contains 1500 pairs of images. These two datasets cover different scenarios, including indoor, outdoor, daytime, and nighttime environments, with each pair of images including a low-light image and a corresponding high-light image. The custom MEG dataset of the present invention includes 1000 groups of images, each group containing 5 images at different exposure levels. To accommodate different training and testing needs, 1000 groups of images with post-processing exposure adjustment are provided, as well as 1000 groups of images taken at different exposure levels through an in-camera rig.

[0113] It is worth noting that the post-processed images maintain perfect content alignment, while the photos in the brackets exhibit slight content changes during capture due to environmental fluctuations. This dataset covers multiple university campuses, different urban scenes, and various weather conditions.

[0114] The training set herein is constructed from 40% of the images randomly selected from the three datasets described above, resulting in a training set containing 1200 pairs of low-light / normal-light image pairs. For the MEG dataset, images are specifically paired with the second and fourth exposure levels to create data pairs. From the remaining 40% of images, normal-light images are selected as style references and randomly matched for the style transfer process. The remaining 20% of images are used as a validation set.

[0115] The evaluation metrics in the experiment are as follows: BRISQUE: measures image distortion and artifacts; NIQE: assesses the naturalness of generated images and their compliance with human visual perception; PIQE: assesses block-based artifacts and noise in images; Gram Metric: quantifies the style similarity between generated images and reference style images using the Gram matrix; Color Histogram Similarity: calculates the similarity between the color distribution histogram of generated images and the style reference image. PSNR: quantifies the peak signal-to-noise ratio between generated images and reference images. SSIM: measures structural similarity between images by comparing brightness, contrast, and structure. LPIPS: uses a deep neural network to assess perceptual similarity, consistent with human judgment.

[0116] Workflow Comparative experiment. Figure 2 The enhancement effect obtained by applying normal-light images with different color tones to low-light inputs at a transformation intensity of 100% is shown. Figure 3The local low-light enhancement effect achieved by single command control is demonstrated. The enhanced results of multiple intensity levels (20%, 40%, 60%, 80%) are visually compared, revealing progressive lighting improvement while preserving the structural details of the shadowed areas.

[0117] First, a series of representative images are tested. Not limited to low-light enhancement scenarios, the evaluation of the present invention covers a variety of brightness levels, color tones, saturation values, and various light characteristics including sunroof, water reflection, urban lighting, and crystal refraction. This comprehensive testing framework thoroughly validates the style transfer performance of the algorithm of the present invention.

[0118] Figure 4 Color transfer results compared with style transfer methods are demonstrated. From Figure 4 It can be observed that the proposed method generally demonstrates robust performance without showing obvious artifacts or excessive noise. Moreover, it maintains reasonable color distribution characteristics after style transfer. To further validate the effectiveness of the algorithm in low-light enhancement, several typical scenes are selected, including sunny, overcast, rainy, night, backlight, and high saturation conditions. Comparative experiments are conducted using various style transfer algorithms for low-light enhancement.

[0119] For low-light images with sunlight scene characteristics, Modflow, CAP-VSTnet, WCT2, and TCSTNet (the present invention) all exhibit good foreground LOL-V2 performance. However, other comparative methods exhibit obvious artifacts in the darkest foreground regions. In terms of sky region processing, CAP-VSTnet results show square abnormal textures, while WCT2 output contains obvious noise.

[0120] Low-light enhancement and style transfer methods are compared under cloudy scenes. Under cloudy conditions, the overall image presents high gray value. Obvious artifacts are observed in both large areas of PhotoWCT and PCA-KD results, while the DeepPreset method cannot maintain effective LOL-V2 capability under low-light conditions. Although other comparative methods exhibit relatively better performance, careful examination finds that WCT2 and Modflow outputs maintain a high gray level. CAP-VSTnet produces superior color tone reproduction but introduces a noticeable blue tone in the shadowed forest. Although the TCSTNet method of the present invention generally achieves satisfactory processing results, slight color bleeding is observed in the red flower area.

[0121] Comparative example data under rainy conditions are as shown in Figure 5The input image is shown in Fig. 6. It can be observed that, due to the higher exposure rate of this input image compared to the previous two groups, no obvious artifacts appear in the results of the algorithm. However, by observing the ground reflection, it can be seen that there is color distortion in the reflection of red light in the PCA-KD method.

[0122] Figure 6 The experimental results under extreme night conditions. As shown in Fig. 8, DeepPreset basically cannot work effectively, while CAP-VSTnet shows local overexposure in the road and sky areas, accompanied by a large number of square noise artifacts. At the same time, WCT2, phoptowct and PCA-KD all produce extensive artifacts in the dark area. Figure 6

[0123] It is worth noting that only the present invention and Modflow achieve satisfactory results. However, since the method of the present invention has a special detail color extraction module, it successfully recovers more accurate color information in key areas such as lane markings and road signs, showing excellent performance in color restoration. Under backlit conditions, the present invention still shows strong performance, as shown in Fig. 9. Specifically, this method neither produces block artifacts nor produces abnormal brightness artifacts observed in the building-sky boundary in the results of CAP-VSTnet and PCA-KD. Figure 7

[0124] Similarly, for images with high saturation, the present invention effectively avoids the generation of block artifacts while demonstrating superior color detail extraction capability. Detailed quantitative indicators such as the Histo indicator are listed in Table 1, which gives the verification results on three different data sets. In addition to comparing with style transfer methods, the present invention also performs comparative analysis with a series of low-light enhancement techniques. To ensure a fair comparison, when running the algorithm of the present invention, the low-light image is used as the content image, and its corresponding well-illuminated version is used as the style image. Then the results are compared with the results obtained by other methods.

[0125] Table 1: Comparative experiments of different style transfer methods

[0126] Four scenarios are selected to demonstrate the results: sunny, rainy, cloudy and indoor conditions, as shown in Fig. 10. The quantitative results on the MEG and LOL data sets are shown in Table 2. Figure 8 to Figure 10

[0127] Table 2: Comparative experiments of different low-light enhancement methods

[0128] Ablation study ​​​Ablation studies were performed on the lol-v2 dataset. The function of each sub-module in the TCSTNet framework of the present application was verified and the influence of adjustable parameters on the algorithm was studied. Ablation studies were performed on the following parts: style loss, super-resolution color extraction module, and encoding step in rectification. The results of the proposed super-resolution module are shown in Figure 11 The visualization results of the ablation study are shown in Figure 12

[0129] The results show that each designed module makes a significant contribution to the overall performance. On the encoding step, increasing the number of steps by 22 generally improves the results, and a decrease in yield of more than 15-25 steps is observed. This trade-off between performance gain and computational efficiency indicates that for most cases, the optimal range of 15-25 encoding steps is sufficient, as further increases: greatly extend the running time, only produce marginal quality improvements. The quantitative results of the ablation study are shown in Table 3.

[0130] Table 3 Ablation study

[0131] In addition, on the dark light segmentation task, the night street view segmentation mIOU (mean intersection over union) on the public dataset Dark Zurich is 64.8%, as shown in Table 4.

[0132] Table 4 Night street view segmentation performance comparison on dataset Dark Zurich

[0133] Other applications In addition to low-light enhancement, the method of the present application has been widely verified in a number of challenging scenarios, including artifact coloring, artistic style transfer, infrared-to-visible spectrum conversion, and underwater image enhancement, showing good performance.

[0134] The advantages of the present application compared with existing methods are shown in Table 5: Table 5 Performance comparison of the present application and existing methods

[0135] 1. High input flexibility: the present application supports multiple inputs such as text-driven instructions, reference style images, etc., rather than being limited to fixed structure or bubble chart input. Users can use natural language descriptions to guide local low-light enhancement.

[0136] ​2. Provide high-fidelity, artifact-free color style transfer: The present application uses rectified flow theory to simulate the color distribution flow conversion from low light to well-exposed images, establishing a continuous mapping between images. Combined with VGG-based style loss and modified flow encoding loss, it ensures that the RGB distribution of the output image is closely consistent with the style image, improving realism and consistency.

[0137] 3. Excellent detail preservation and noise suppression capability: The present application designs a super-resolution module combining U-Net, bilateral denoising module and residual convolutional neural network (CNN) to accurately extract color details from the original image. The bilateral filter effectively reduces down-sampling noise while maintaining edge sharpness.

[0138] In summary, the present application proposes a language feature fusion image color migration network low light image enhancement method for the complex and fuzzy mapping relationship between low light images and well-exposed images in LLIE, and the situation that traditional color style transfer technology is prone to artifacts, low image fidelity, and limited explainability. A text-driven color style transfer network (TCST-N) is constructed.

[0139] The method of the present application is based on text-driven and rectified flow theory, which can effectively realize personalized and semantic-aware image enhancement by fully utilizing natural language description to guide local low light enhancement and global color style transfer, and establishing the characteristics of color distribution flow conversion relationship. This is due to its treatment of images as RGB space probability distribution, smooth transmission of color distribution through ordinary differential equation and linear interpolation, and precise local area control through LISA model.

[0140] The method of the present application is based on Rnx theory and deep learning, which can effectively realize more accurate and detailed color extraction and noise suppression, and improve the realism and consistency of style transfer by fully utilizing U-N, bilateral denoising module and residual convolutional neural network for color detail extraction, and the characteristics of VGG-basd style loss function and modified flow encoding loss. In addition, by constructing a multi-exposure gradient (MEG) dataset and using a modular training paradigm and composite loss function, the superior performance and generalization ability of the method in multiple challenging scenarios are verified.

[0141] Important features of the invention include: deep fusion of low-light image enhancement (LLIE) with color style transfer (CST) rather than independent processing. Using rectified flow theory, modeling the color distribution flow conversion from low-light images to well-exposed images, establishing continuous mapping between images. In the encoder, using Retinex theory to decompose low-light images into illumination and reflectance components for independent color information extraction and brightness detail enhancement. The entire system is based on deep convolutional neural networks (CNNs), U-Net, and other deep learning architectures for image feature extraction, conversion, and generation.

[0142] Alternative embodiments: 3D scene reconstruction and stylization: extending rectified flow theory and semantic segmentation to 3D point cloud or mesh data, achieving lighting reconstruction and artistic style transfer for low-light 3D scenes.

[0143] Reverse style transfer and style learning: developing inverse models that allow users to extract the underlying textual descriptions or style parameters from enhanced or stylized images for new creation or personalized recommendations.

[0144] Interactive editing of style transfer: developing a user interface that allows users to make real-time, fine-grained local adjustments to the effects of style transfer through brushes, region selection, etc. based on textual instructions.

[0145] Integrating embodied intelligence for scene perception enhancement: integrating the model into robots or AR / VR devices to enable autonomous adjustment of visual output based on real-world environmental lighting conditions and semantic information, providing a more immersive and adaptive experience.

[0146] The invention has input modalities: in addition to textual instructions, other modalities (such as audio, hand-drawn sketches, emotional labels) can be introduced to drive style transfer.

[0147] Platform and form of model deployment: cloud services, local software, SDK / API for embedded devices (such as smartphones, surveillance cameras), etc. different deployment forms.

[0148] Applications The method of the invention has a wide range of application scenarios. Specific applications include but are not limited to:

[0149] Monitoring and security field: TCSTN can significantly enhance the visibility, contrast and perceptual quality of surveillance images while preserving natural details and color fidelity.

[0150] E-commerce field: TCSTN allows users to achieve local low-light enhancement and global lighting adjustment through natural language instructions, providing highly customized experiences for users.

[0151] The field of cultural relics digitization and the like: greatly reducing the dependence on hardware devices and manual operation for obtaining low-light images.

[0152] The above are only examples of application fields, and the application of the text-driven color style transfer network in practice can also involve more fields.

[0153] The embodiment of the present application also provides a storage medium for storing a computer program, which is executed to perform at least the method described above.

[0154] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent substitutions or obvious modifications can be made, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present application.

Claims

1. A low-light image enhancement method based on an image color transfer network integrating language features, characterized in that: The following steps are involved: S1. Color Distribution Modeling and Transfer: Based on rectification theory, the input low-light content image and reference style image are modeled as three-dimensional probability distributions in RGB space. The transfer process between the distributions is described by ordinary differential equations. Linear interpolation is used to construct intermediate states and fit the transfer mapping function to achieve smooth migration of color distribution from the content image to the style image. S2. Super-resolution color feature extraction: Based on the Retinex theory, low-light images are decomposed into illumination and reflectance components. A dual-branch encoder is used to extract brightness details and color information respectively. The reflectance branch integrates a U-Net structure and a bilateral filter for noise suppression and edge preservation. The illumination branch uses a residual convolutional network to enhance structural details. Super-resolution images are generated through cross-branch feature fusion. S3, Text-driven region enhancement: Parses user text instructions and decomposes region descriptors and intensity coefficients using a language model. It combines the reflectivity component and region descriptor to generate a spatially accurate enhancement mask. It controls the rectification transmission process based on the mask and intensity coefficients to achieve iterative optimization and adaptive mixing of local regions. S4. Modular joint training: Pre-train the Retinex decomposition module, super-resolution module, and rectifier encoder in stages, and perform end-to-end fine-tuning through a composite loss function to ensure that the color distribution and structural details of the output image are consistent with the style reference.

2. The method according to claim 1, wherein Step S1 specifically includes the following process: The pixel samples of the content image and style image are regarded as corresponding random variables respectively; The intermediate state variables between the two are constructed through linear interpolation in the time dimension; Define the ordinary differential equation that describes the distribution transfer process and optimize the transfer function by minimizing the fitting loss between the transfer function and the intermediate state variables; A latent distribution that obeys uniform distribution on a uniform interval is introduced to bridge the content image distribution and the style image distribution. Through the combination of forward sampling of the content image distribution and backward sampling of the style image distribution, the color distribution transmission from the content image to the style image is achieved.

3. The method according to claim 2, wherein The transfer mapping and its inverse mapping are realized by integrating the transfer function: The forward transfer map is obtained by forward integrating the transfer function along the time domain, which is used to transfer from the content image distribution to the intermediate latent distribution; The backward transfer mapping is obtained by integrating the transfer function backward along the time domain, which is used to transfer from the intermediate latent distribution to the style image distribution.

4. The method according to any one of claims 1 to 3, wherein In step S2, the specific operation process of the dual-branch encoder includes: Reflectivity branch: A U-Net structure is used for downsampling and upsampling operations. An edge-preserving denoising module is inserted between each downsampling layer and the corresponding upsampling layer to suppress the noise introduced during the downsampling process. Illumination branch: A residual network consisting of multiple layers of convolution and activation functions is used to extract high-frequency brightness detail features from the illumination image. In the first and last layers of U-Net, an element-by-element multiplication operation between the output of the illumination branch and the output of the reflectance branch is introduced, so that the reflectance branch can adaptively focus on the chromaticity information; The outputs of the two branches are fused through element-wise multiplication, and then weighted summed with the input image, and a super-resolution image is generated after tail processing.

5. The method according to claim 4, wherein The edge-preserving denoising module uses a bilateral filter to bridge each downsampling-upsampling pair of the U-Net, suppressing the noise introduced during the downsampling process while maintaining the clarity of the image edges.

6. The method according to any one of claims 1 to 3, wherein: In step S2, the super-resolution image is mapped to weight parameters of a rectification transfer function through a convolutional neural network, and an association between super-resolution image features and the rectification transfer process is established.

7. The method according to any one of claims 1 to 3, characterized in that Step S3 specifically includes the following processing steps: The language model is used to parse the text instructions input by the user, and output intensity parameters for regulating the degree of enhancement and semantic region descriptions for locating the target area; The segmentation model combines the reflectivity component obtained from the Retinex encoder to generate a spatially accurate binary mask based on the semantic region description to identify the local area that needs to be enhanced; For the area covered by the mask, a rectification transfer process under the control of intensity parameters is applied to achieve style transfer and enhancement in this area; By iteratively optimizing the rectification transmission parameters, the enhancement effect at the mask boundary is made to transition smoothly, avoiding obvious regional segmentation traces.

8. The method according to any one of claims 1 to 3, wherein: The staged pre-training in step S4 specifically includes the following stages that are performed in sequence: The first stage: with the decomposition loss of reflectance and illumination components as the target, a dual-branch encoder is trained so that the decomposed reflectance and illumination components can accurately reconstruct the input low-light image; The second stage: The training image is downsampled and then upscaled to its original size to simulate blurred input. The super-resolution module is trained with a multi-constrained reconstruction loss, enabling the module to reconstruct a high-resolution image from the blurred input. The third stage: The pre-trained super-resolution module is combined with the rectifier encoder to perform end-to-end fine-tuning with the goal of integrating a composite loss function that integrates multiple losses to optimize the overall network parameters.

9. The method according to any one of claims 1 to 3, wherein: The composite loss function specifically includes the following components: Coding loss: ensures the fitting accuracy of the transfer function by constraining the difference between the predicted value and the true value of the transfer function; Style loss: Using features extracted by deep neural networks, we match the deep features of the generated image with the reference style image to constrain the consistency of the two styles. Content loss: By constraining the pixel-level difference between the generated image and the content image, the structural and content consistency of the image is maintained; Reconstruction loss: Integrates constraints on image gradient details, regularization constraints on color space, and measurement constraints on structural similarity to ensure the details and quality of the generated image.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for enhancing low-light images by image color migration network integrating language features as described in any one of claims 1 to 9 is implemented.

Images