Low-light image enhancement and adaptive dimming method based on semi-supervision

Abstract

The invention discloses a low-light image enhancement and adaptive dimming method based on semi-supervision, and the method comprises the following steps: constructing a dimming module based on a mixed density network, and the dimming module is used for simulating the color distortion and darkening process under a low-light condition; a dimming module is used for converting the high-quality image into a low-illumination image, and a synthesized low-light-normal-light image pair is generated to serve as training data; training an image enhancement model of a UNet architecture on the synthetic training data in an unsupervised manner, and performing fine tuning by using a small number of real images; in the image enhancement process, the low illumination effect is controlled by setting a dimming factor, and the brightness level of an output image is controlled through condition input. The method has the beneficial effects that the technical scheme realizes multi-dimensional breakthrough in the aspects of data efficiency, reality sense simulation and enhancement quality and flexibility through a semi-supervised learning strategy, a physically-driven dimming module and adaptive brightness control, and effectively solves the core pain point of a traditional low-light enhancement method.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing, and in particular to a semi-supervised low-light image enhancement and adaptive dimming method. Background Technology

[0002] Low-light image enhancement is a key technology in computer vision and image processing, widely used in scenarios such as nighttime surveillance, mobile phone photography, and autonomous driving visual perception. Its core challenge lies in enhancing image brightness while preserving detail, suppressing noise, and maintaining natural color distribution in the original scene. However, existing technologies still have the following significant limitations in practical applications: Traditional fully supervised methods rely on a large amount of paired low-light to normal-light image data (such as the LOL dataset) for training. While the models perform well on specific datasets, their generalization ability drops significantly when faced with different camera devices, scenes, or lighting conditions due to differences in imaging characteristics (such as sensor noise and lens distortion). Unsupervised methods, on the other hand, eliminate the dependence on paired data and achieve enhancement through image decomposition (such as Retinex theory) or generative adversarial networks (GANs). However, they often suffer from problems such as overexposure, color distortion, or blurred details, making it difficult to balance brightness enhancement with naturalness.

[0003] Existing methods are mostly trained on synthetic low-light data, which differs from real low-light environments in terms of distribution. For example, synthetic data often ignores the nonlinear response of camera hardware in low light (such as noise caused by ISO gain and white balance shift), making it easy for models to produce color casts or artifacts when processing real low-light images. In addition, the differences in low-light imaging characteristics of different brands and models of cameras further exacerbate the difficulty of model adaptation.

[0004] Most enhancement algorithms employ a "fixed target brightness" strategy, which fails to meet users' personalized brightness requirements for the output image (such as adjusting brightness according to the scene). Some methods attempt to introduce brightness adjustment parameters, but these are mostly implemented through post-processing, which can easily disrupt the overall contrast and color consistency of the image.

[0005] Although semi-supervised learning has shown potential in reducing data dependence, the design of semi-supervised strategies in existing low-light enhancement methods is still immature: they either simply mix a small amount of real data with synthetic data without modeling the physical characteristics of low-light imaging; or they lack refined processing of key factors such as reflectivity and illumination components, resulting in enhancement effects that are inferior to fully supervised methods.

[0006] To address the aforementioned issues, this invention proposes a semi-supervised method that combines a hybrid density network dimming module with a conditional UNet architecture. This method aims to achieve a dual breakthrough in data efficiency and enhanced quality by simulating the real low-light imaging process and dynamically controlling the brightness output. Summary of the Invention

[0007] The main technical problem solved by this invention is to provide a semi-supervised low-light image enhancement and adaptive dimming method, thereby solving one or more of the problems mentioned above in the prior art.

[0008] To solve the above-mentioned technical problems, the present invention adopts a technical solution as follows: a semi-supervised low-light image enhancement and adaptive dimming method, the innovation of which is: including the following steps: S1: Construct a dimming module based on a hybrid density network, which is used to simulate color distortion and darkening processes under low light conditions; S2: Use the dimming module to convert the high-quality image into a low-light image, and generate a synthesized low-light-normal-light image pair as training data; S3: Train the UNet architecture image augmentation model on the synthetic training data in an unsupervised manner, and fine-tune it using a small number of real image pairs; S4: During image enhancement, the low-light effect is controlled by setting the dimming factor, and the brightness level of the output image is controlled by the conditional input.

[0009] In some implementations, in step S1, the dimming module calculates the reflectivity R of the i-th channel by channel-by-channel normalization, using the following formula: in, and For the input image and the first of the reflectivity R aisle, This indicates Hadamard element-level partitioning.

[0010] In some implementations, in step S1, the dimming module adjusts the reflection component R through a hybrid density network. The MDN outputs the adjusted reflection component RD based on the input reflection component R, normal illumination component L, and dimming illumination component LD, and combines RD and LD through a Hadamard product to obtain the dark image ID = RD ⊙ LD.

[0011] In some implementations, in step S2, the high-quality image comes from a publicly available computer vision dataset, and the synthesized image generated by the dimming module is used for the unsupervised training phase of the model.

[0012] In some implementations, in step S3, the fine-tuning process uses real image pairs from when the dimming module was built to adapt to the imaging characteristics of a particular camera under low-light conditions.

[0013] In some implementations, the dimming factor is a γ value, and different levels of low-light effects are generated by adjusting γ; the conditional input is a brightness value Δm, which is embedded in the UNet architecture to control the brightness level of the output image.

[0014] In some implementations, the MDN processes the 5-dimensional representation of each pixel independently, shares weights through a multilayer perceptron, and outputs the mean, standard deviation, and mixing coefficient of each color channel to describe the probability distribution of color distortion.

[0015] In some implementations, the embedding of the brightness value Δm is similar to the temporal embedding in the diffusion model, allowing UNet to dynamically adapt to the target brightness during the enhancement process.

[0016] In some implementations, in step S3, the input of the UNet model includes a darkened image, a darkened image after histogram equalization, color mapping and illumination information, and outputs a residual image U. The final enhanced image is obtained by adding the residual image U to the darkened image.

[0017] In some implementations, the semi-supervised approach is achieved through the following steps: A large number of synthetic low-light-normal-light image pairs are generated using a dimming module, and unsupervised pre-training is performed. The UNet model was fine-tuned using a small number of real low-light to normal-light images to balance generalization ability with scene-specific adaptability.

[0018] The beneficial effects of this invention are as follows: This technical solution reduces the dependence on large-scale paired low-light-normal-light image data by adopting a semi-supervised strategy of "unsupervised pre-training with synthetic data + fine-tuning with a small amount of real data"; it completes the initial training by using a large amount of synthetic data (from public datasets) generated by the dimming module, and then fine-tunes it with a small amount of real image pairs (such as real camera shooting data), which not only solves the problem of scarcity of real data, but also improves the model's adaptability to different camera imaging characteristics.

[0019] This technical solution utilizes diverse public datasets in the unsupervised pre-training stage, enabling the model to learn general low-light enhancement rules. The fine-tuning stage optimizes for specific camera characteristics, balancing "generality" and "scene specificity," effectively alleviating the contradiction between the poor generalization ability of traditional fully supervised methods and the limited effectiveness of unsupervised methods. Based on Retinex theory, this solution decomposes images into reflectance (R) and illumination (L) components. Reflectance is calculated through channel-by-channel normalization, and a hybrid density network (MDN) is introduced to model the color distortion probability distribution. The MDN independently processes the 5-dimensional features (reflectance + illumination conditions) of each pixel, outputting the mean, standard deviation, and mixing coefficients to accurately simulate color shift and darkening processes in real low-light environments. The generated synthetic low-light images have a distribution consistent with real-world images. In the dimming module training stage, this solution uses a small number of real images to learn the low-light imaging characteristics of specific cameras (such as sensor noise and white balance shift), ensuring that the synthetic data includes device-related characteristics and solving the "distribution offset" problem between traditional synthetic data and real-world scenes. This technical solution generates different levels of low-light effects by setting a dimming factor γ, supporting the construction of diverse training data and meeting users' personalized needs for low light intensity. The solution embeds a brightness value Δm as a conditional input into the UNet architecture, allowing users to directly specify the target brightness level. The model dynamically adjusts the enhancement intensity through residual learning, avoiding contrast imbalance or color cast caused by traditional post-processing adjustments, achieving a "controllable and natural" enhancement effect. The UNet model input in this solution integrates a darkened image, a histogram-equalized image, color mapping, and lighting information. This rich input modality provides the model with global contrast and local detail cues, improving the ability to recover details in dark areas. The model outputs a residual map U, and the final enhancement result is obtained by combining the "darkened image + residual map," avoiding overexposure or noise amplification problems that easily occur when directly generating images, ensuring that image naturalness is maintained while increasing brightness.

[0020] This technical solution does not rely on professional high dynamic range imaging equipment. It achieves low-light enhancement of ordinary cameras through algorithm optimization, thus lowering the application threshold. It can be applied to fields such as night monitoring, mobile phone photography, and autonomous driving visual perception, and is especially suitable for scenarios with high requirements for real-time performance and robustness.

[0021] This technical solution achieves breakthroughs in data efficiency, realistic simulation, enhancement quality, and flexibility through a semi-supervised learning strategy, a physics-driven dimming module, and adaptive brightness control, effectively addressing the core pain points of traditional low-light enhancement methods. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a diagram illustrating the process of training the UNet model using the dimming module according to an embodiment of the present invention.

[0023] Figure 2 The process of creating a corresponding dark image for the dimming process described in the embodiments of the present invention.

[0024] Figure 3 This is a diagram illustrating the architecture of the MDN network described in an embodiment of the present invention.

[0025] Figure 4 This is a structural diagram of the brightness module according to an embodiment of the present invention.

[0026] Figure 5 Visual comparison of Dimma image pairs from different fully supervised methods in the LOLdata set. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] The present invention includes a semi-supervised low-light image enhancement and adaptive dimming method that enhances low-light images by generating synthetic data through a dimming module, unsupervised pre-training, and fine-tuning with a small amount of real data, and supports user adaptive control of output brightness. Its core steps include: constructing a dimming module based on a hybrid density network (MDN); generating synthetic low-light-normal light image pairs; training a UNet enhancement model with conditional input; and adaptively adjusting the output image brightness.

[0029] S1: Constructing a dimming module based on a hybrid density network The core function of this step is to simulate the color distortion and darkening process of a specific camera under low light conditions, decompose the image into a reflection component (R) and an illumination component (L) using Retinex theory, and adjust the reflection component using MDN to generate a realistic low-light image.

[0030] Reflectance calculation: The reflectance R of the i-th channel is calculated by normalizing each channel. The calculation formula is as follows: in, and For the input image and the first of the reflectivity R aisle, This indicates Hadamard element-level partitioning.

[0031] Hybrid density network (MDN) adjusts the reflection component: MDN Inputs: Reflection Component R, Normal Illumination Component L, Dimming Illumination Component ; MDN Output: Adjusted Reflection Components (The color distortion probability distribution is modeled using a multilayer perceptron (MLP), and the mean, standard deviation, and mixing coefficient of each channel are output.) Generating dark images: Combining via Hadamard product and To obtain a dark image = ⊙ (⊙ represents Hadamard element-wise multiplication).

[0032] S2: Generate a synthetic low-light-normal-light image pair Data source: High-quality normal light images from publicly available computer vision datasets (such as LOL dataset, MSCOCO, etc.) are used as raw data.

[0033] Synthesis process: The dimming module constructed in step S1 converts the normal light image into low-light images of varying degrees, forming a synthesized low-light image-normal light image pair; specifically including: Set the dimming factor γ to control the level of low light (the larger the γ value, the darker the image); Generate a large amount of synthetic data to meet the needs of unsupervised training.

[0034] S3: Training an image augmentation model using the UNet architecture Unsupervised pre-training: like Figure 1 and Figure 2 As shown, the UNet model is trained in an unsupervised manner on the synthetic dataset generated in step S2; the UNet input includes: darkened image, histogram equalized darkened image, color mapping and illumination information; Model output residual plot Ultimately, the image is enhanced through calculate.

[0035] Fine-tuning based on real data: The UNet model is fine-tuned using a small number of real low-light-normal-light image pairs (such as real-world camera data) to adapt to the low-light imaging characteristics of specific cameras, balancing generalization ability and scene adaptability.

[0036] S4: Adaptive dimming and brightness control Dimming factor adjustment: By adjusting the dimming factor γ, images with different low light intensities are generated (e.g., γ=0.3 corresponds to mild low light, γ=0.8 corresponds to severe low light).

[0037] Conditional input controls brightness: In the UNet architecture, a brightness value Δm is embedded as a conditional input (similar to the temporal embedding in the diffusion model), and the user can control the output image brightness by specifying Δm. For example: =0 corresponds to the original brightness, Δm=+0.2 increases the brightness, and Δm=−0.1 decreases the brightness.

[0038] Key module design like Figure 3 As shown, the MDN network structure is as follows: Input: Pixel-level 5D features (reflectivity R, normal illumination L, dimmed illumination) ); Output: Gaussian distribution parameters (mean, standard deviation, mixing coefficient) for each color channel, used to model the probability distribution of color distortion.

[0039] like Figure 4 As shown, the brightness control module: Conditional input: User-specified brightness value ; Working mechanism: Through the embedding layer Integration The feature extraction process dynamically adjusts and enhances the brightness level of the image.

[0040] IV. Experimental Verification like Figure 4 As shown, compared with existing fully supervised methods on the LOL dataset, the present invention performs better in terms of brightness enhancement, color fidelity and detail preservation, especially in terms of stronger generalization ability on low-light images taken by different cameras.

[0041] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A semi-supervised low-light image enhancement and adaptive dimming method, characterized in that: Includes the following steps: S1: Construct a dimming module based on a hybrid density network, which is used to simulate color distortion and darkening processes under low light conditions; S2: Use the dimming module to convert the high-quality image into a low-light image, and generate a synthesized low-light-normal-light image pair as training data; S3: Train the UNet architecture image augmentation model on the synthetic training data in an unsupervised manner, and fine-tune it using a small number of real image pairs; S4: During image enhancement, the low-light effect is controlled by setting the dimming factor, and the brightness level of the output image is controlled by conditional input.

2. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: In step S1, the dimming module calculates the reflectivity R of the i-th channel by normalizing it channel by channel. The calculation formula is: in, and For the input image and the first of the reflectivity R aisle, This indicates Hadamard element-level partitioning.

3. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: In step S1, the dimming module adjusts the reflection component R through a hybrid density network. The MDN outputs the adjusted reflection component RD based on the input reflection component R, normal illumination component L, and darkened illumination component LD. The dark image ID = RD⊙LD is obtained by combining RD and LD through a Hadamard product.

4. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: In step S2, the high-quality images are from publicly available computer vision datasets, and the synthetic images generated by the dimming module are used for the unsupervised training phase of the model.

5. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: In step S3, the fine-tuning process uses real image pairs from when the dimming module was built to adapt to the imaging characteristics of a specific camera under low-light conditions.

6. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: The dimming factor is a γ value, which generates different levels of low-light effects by adjusting γ; the conditional input is a brightness value Δm, which is embedded in the UNet architecture to control the brightness level of the output image.

7. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: The MDN processes the 5-dimensional representation of each pixel independently, shares weights through a multilayer perceptron, and outputs the mean, standard deviation, and mixing coefficient of each color channel to describe the probability distribution of color distortion.

8. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: The embedding method of the brightness value Δm is similar to the temporal embedding in the diffusion model, enabling UNet to dynamically adapt to the target brightness during the enhancement process.

9. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: In step S3, the input of the UNet model includes the darkened image, the darkened image after histogram equalization, color mapping and illumination information, and the output is the residual image U. The final enhanced image is obtained by adding the residual image U to the darkened image.

10. The semi-supervised low-light image enhancement and adaptive dimming method according to claim 1, characterized in that: The semi-supervised method is achieved through the following steps: A large number of synthetic low-light-normal-light image pairs are generated using a dimming module, and unsupervised pre-training is performed. The UNet model was fine-tuned using a small number of real low-light to normal-light images to balance generalization ability with scene-specific adaptability.

Images