Low-illumination image enhancement method and system based on deep learning

Abstract

The invention discloses a low-illumination image enhancement method and system based on deep learning, and belongs to the field of image enhancement, and the image enhancement method comprises the following steps: S1, obtaining a remote sensing image: the remote sensing image comprises a pairing data set of indoor low-light and normal-light images; s2, performing feature extraction, filtering and denoising on the acquired remote sensing image; s3, constructing an image enhancement model based on the improved Transform network; s4, designing a loss function and setting different training parameters, and training the model to obtain an image enhancement model; s5, inputting the test set low-illumination image into the image enhancement model for testing; s6, obtaining an enhanced image; the invention also discloses a low-illumination image enhancement system based on deep learning.

Description

Technical Field

[0001] This application belongs to the field of computer vision and image processing technology, and specifically relates to a low-light image enhancement system and method based on deep learning, which is suitable for optimizing image quality in low-light scenarios such as nighttime monitoring, autonomous driving, medical imaging, and electronic equipment. Background Art

[0002] With the advancement of digitalization, digital images have become a core medium for visual information. However, complex lighting conditions (such as night scenes and backlighting) can lead to image artifacts such as low illumination, high noise, and color distortion, which impact perception and subsequent visual tasks (such as object detection). Therefore, to optimize the viewing experience and enhance information discernibility, image enhancement technologies for low-light applications have rapidly developed. Traditional enhancement methods (such as histogram equalization and Retinex theory) rely on artificial feature design and suffer from drawbacks such as poor scene adaptability and noise amplification. While existing deep learning-based solutions (such as EnlightenGAN and Zero-DCE) can improve overall brightness, they still suffer from localized uneven illumination, residual noise, and imbalanced metrics. Furthermore, they face bottlenecks such as insufficient model generalization, blurred edges, and high computational complexity. This study systematically analyzes existing methods and proposes an innovative design that optimizes noise suppression, multi-metric balancing, and exposure uniformity, thereby enhancing the practical application value of the enhancement results. Summary of the Invention

[0003] To solve the above problems, this patent proposes a low-light image enhancement method based on deep learning. The invention provides the following technical solutions and steps: S1. Organized dataset: LOL, Exclusive Dark, a paired dataset containing low-light / normal-light images.

[0004] S2. Perform preprocessing operations such as feature extraction, filtering, and denoising on the images in the dataset; S3. Build an image enhancement model based on the improved Transformer network during the model training phase; S4. Design a loss function and set different training parameters to train the model and obtain the optimal model; S5. Input the low-light images in the test set into the image enhancement model for testing; S6. Obtain the enhanced image.

[0005] Step S2 proposes an RCBAM attention mechanism, which is divided into a channel attention module and a spatial attention module. It also designs a path to preserve the feature information of the original image. The purpose is to extract high-level feature information when the image passes through the backbone network. After passing through the RCBAM attention module, the feature information of the low-light areas in the image is highlighted, and the low-light local features are emphasized and exposed, while the original brighter areas are slightly enhanced to prevent overexposure of the brighter areas of the original image, which may cause image quality problems.

[0006] This solution uses the channel attention mechanism to optimize low-light image features. The core process is as follows: Dual pooling spatial compression: Max pooling (MaxPool) and average pooling (AvgPool) are performed in parallel on the input feature map (C×H×W) to generate two channel feature vectors (C×1×1), preserving key feature information; Shared MLP feature interaction: The double pooling results are input into a shared multi-layer perceptron (MLP), which performs dimensionality reduction (compression rate 1 / r), ReLU activation, and dimensionality increase in sequence to extract nonlinear correlation feature information between channels; Feature fusion and weight allocation: The two MLP outputs are added channel by channel, and the channel weights (0-1) are generated by the Sigmoid function. This dynamically strengthens important channel features, suppresses noise interference, and effectively guides which features should be enhanced or weakened in subsequent processing, thereby greatly improving the pertinence and efficiency of image processing. The formula is as follows: Spatial attention module design, this module achieves spatial feature optimization through two-level attention collaboration: Input processing: The weighted feature maps output by Channel Attention are MaxPooled and AvgPooled in the channel dimension to generate two 1×H×W spatial feature maps, retaining the key area response; Feature fusion: The double pooling results are combined through the concatenation operation to form a 2×H×W composite feature map, fusing local extreme values ​​with global statistical information; Spatial weight generation: A 7×7 convolution kernel is used to extract wide-area contextual relationships, which are compressed into a single-channel feature map and then activated with a sigmoid to generate a spatial attention heat map (1×H×W) to quantify the weight information of each pixel region. Feature recalibration: The spatial attention map is multiplied point by point with the original C×H×W feature map to dynamically enhance the feature representation of the target region. This process strengthens the representation of the target region. The final output is still a C*H*W feature map, but the feature values ​​at each location have been carefully adjusted to make it more suitable for subsequent visual processing tasks.

[0007] Step S2 proposes a denoising module and a noise estimation module architecture: a 5-layer fully convolutional network (FCN) generates a noise level map and accurately models the spatial noise distribution.

[0008] Optimization: Add a parallel 5-layer FCN branch to dynamically learn and optimize the noise intensity parameter (σ) input by the user to achieve adaptive adjustment.

[0009] Downsampling and parallel processing strategy: Divide the image into four sub-images to reduce computational complexity and support parallel processing. Noise mapping: A spatially invariant AWGN noise level map (with a uniform σ value) is used to guide sub-image denoising.

[0010] Denoising and reconstruction algorithms: Based on sub-image characteristics and noise mapping, they specifically suppress noise while preserving details. Sub-images are then reassembled through upsampling and interpolated to restore the full size, avoiding distortion.

[0011] This system uses a five-layer fully convolutional network to achieve intelligent noise estimation and generates accurate noise level maps through deep learning. To improve adaptability, the system introduces adjustable noise intensity parameters and is equipped with a dedicated network branch for parameter optimization, realizing the upgrade from fixed parameters to adaptive strategies. The processing flow adopts a four-equal-division downsampling scheme to divide the image into sub-images for parallel processing, and cooperates with a unified AWGN noise mapping for targeted noise reduction. Finally, the complete image is reconstructed through high-quality upsampling technology, which effectively balances the denoising effect and detail retention while ensuring processing efficiency. The system realizes the intelligent removal of spatially invariant noise through the collaborative design of noise estimation network and adaptive parameters, combined with an efficient parallel processing architecture, providing an efficient and high-quality solution for image processing. The formula is as follows: LCAN loss function design: The loss function is the core mechanism for deep neural network optimization and directly affects model performance. In low-light image enhancement tasks, a comprehensive loss function is designed to improve image visual quality (including illumination uniformity, color restoration, contrast enhancement, and noise suppression). Traditional single loss functions have limitations. This study proposes a multi-component loss framework: Pixel-level error: MSE or SSIM is used to ensure the brightness / structure similarity between the reconstructed image and the real image. The formula is as follows: Light loss: Optimize the global brightness distribution to approach natural lighting conditions. The formula is as follows: Color loss: Enhances color naturalness and distinction, and improves visual layering. The formula is as follows: , Noise suppression loss: Reduce noise and preserve details through TV regularization or adaptive weights. The formula is as follows: Hyperparameter optimization uses grid search, random search, and Bayesian optimization to balance the weights of various loss terms. The overall loss function is shown in the formula. This design can guide the network to improve image quality in multiple dimensions, providing a reference for image processing and other multi-metric optimization tasks. The overall loss function formula is as follows.

[0012] .

[0013] The core of this invention lies in further exploring the field of low-light image enhancement. This begins with an extensive investigation of existing algorithms to systematically analyze and identify the limitations and flaws of existing technologies. From this foundation, we will identify the direction for improvement and devise innovative strategies to fill the gaps in existing approaches.

[0014] The algorithm design phase consists of three parts. The first task is to carefully design the network architecture. This project proposes a novel attention mechanism, RCBAM, to optimize the feature extraction process and ensure that the network can more accurately focus on key features in the image. Secondly, we will improve a loss function that combines illumination, color, contrast, saturation, and spatial position to ensure more precise goal-oriented training and drive the model towards more optimal solutions. Furthermore, we will design a denoising module, called FCBR, to improve image purity and reduce noise interference. Next, we will enter the module construction phase, focusing on constructing an optimized Transformer network. This network integrates the aforementioned design ideas, integrating the attention mechanism, denoising module, and loss function into the backbone network, aiming to efficiently perform low-light image enhancement tasks. Subsequently, experimental comparative analysis will help us fine-tune the network model based on actual test feedback. During this phase, we will iteratively optimize the model based on experimental output to ensure continuous performance improvement. The research then enters the practical verification phase, which is not only a process of translating theory into practice but also a critical period for testing and optimizing research results. First, we developed efficient and maintainable code to build an algorithmic framework that could support the innovative ideas of this project. We then utilized advanced development tools and platforms to meticulously deploy and execute model training tasks.

[0015] During this process, repeated parameter adjustments became the norm. Based on experimental feedback on performance metrics like accuracy, loss function value, and training speed, network hyperparameters such as learning rate, batch size, and regularization strength were constantly fine-tuned, striving to find the optimal balance between network complexity and generalization ability. After each parameter adjustment, the experiment was rerun to observe the impact of the change. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the structure of the low-light image enhancement system based on deep learning of the present invention; Figure 2 Schematic diagram of the RCBAM attention module structure of the low-light image enhancement system based on deep learning of the present invention; Figure 3 Schematic diagram of the FCBD denoising module structure of the low-light image enhancement system based on deep learning of the present invention; Figure 4 This is an experimental rendering of the low-light image enhancement system based on deep learning of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The embodiments of the present invention are further described in detail below in conjunction with the drawings.

[0018] like Figure 1 As shown, the present invention discloses a low-light image enhancement structure diagram based on deep learning, and the method includes the following steps: S1. Organized dataset: LOL, Exclusive Dark, contains paired datasets of indoor low-light / normal-light images.

[0019] S2. Perform preprocessing operations such as feature extraction, filtering, and denoising on the dataset images; S3. Build an image enhancement model based on the improved Transformer network during the model training phase; S4. Design a loss function and set different training parameters to train the model and obtain the optimal model; S5. Input the low-light images in the test set into the image enhancement model for testing; S6. Obtain the enhanced image.

[0020] like Figure 2 As shown in the figure, in the preprocessing stage (S2), the present invention innovatively proposes a dual-path RCBAM (Residual Channel-Block Attention Module) attention mechanism, which consists of three core components: Feature-preserving path: Establishes a direct connection channel to fully preserve the original high-level features extracted by the backbone network. Through residual connections, it ensures that basic image information is not lost, maintains the original brightness distribution characteristics, and prevents information attenuation.

[0021] Intelligent Lighting Adjustment Module Channel Attention Unit: A differentiated feature enhancement strategy is used to enhance the features of low-light areas by 3-5 times, and only a slight enhancement of 1.2-1.5 times is implemented for normal brightness areas.

[0022] The spatial attention unit generates a light-adjustment heatmap based on region importance analysis, dynamically categorizing enhancement intensity levels into weak, medium, and strong regions. The overexposure protection mechanism automatically applies gain suppression (enhancement factor ≤ 1.2) to bright areas above a set luminance threshold (>180 lux). This nonlinear mapping prevents saturation. This design, combining a feature-preserving and attention-enhancing pathway, ensures the integrity of original image features while significantly improving visibility in low-light areas (by 200-300%), moderately optimizing detail in medium-light areas (by 50-80%), and intelligently maintaining a natural appearance in bright areas (by ≤20%). This region-adaptive enhancement strategy effectively resolves the over-enhancement / under-enhancement trade-off common in traditional methods, resulting in an average 4.2dB improvement in PSNR across multiple tests.

[0023] like Figure 3 As shown, a denoising module called FCBD was also designed to improve image purity and reduce noise interference. The module construction phase then began, focusing on constructing an optimized Transformer network. This network incorporates the aforementioned design ideas, integrating the attention mechanism, denoising module, and loss function into the backbone network, aiming to efficiently perform low-light image enhancement tasks. Subsequent experimental comparative analysis helped us fine-tune the network model based on actual test feedback. During this phase, we continuously iterated and optimized the model based on experimental output to ensure continued performance improvement.

[0024] The research then entered the crucial phase of practical validation. This phase is not only about translating theory into practice, but also a crucial period for testing and optimizing research findings. First, efficient and maintainable code was developed to construct an algorithmic framework capable of implementing the innovative ideas of this project. Leveraging advanced development tools and platforms, model training was meticulously deployed and executed. Supervised datasets were applied to this low-light image enhancement algorithm, and the performance parameters of the enhanced images were observed by calculating the loss function. This process involved iterative parameter tuning. Based on experimental feedback from performance metrics such as accuracy, loss function values, and training speed, the network architecture's hyperparameters, such as the learning rate, batch size, and regularization strength, were continuously fine-tuned to achieve the optimal balance between network complexity and generalization ability. After each parameter adjustment, the experiment was rerun to observe the impact of the change.

[0025] All experiments in this paper were conducted using the PyTorch 1.3.1 deep learning framework on a 64-bit Windows 10 operating system, equipped with an NVIDIA GeForce RTX 3080 graphics card, 8GB of RAM, and accelerated by CUDA 11.8. The model was trained for 300 iterations.

[0026] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0027] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. A low-light image enhancement method based on deep learning, characterized in that: Here are the steps: S1. Acquire remote sensing images: The remote sensing images include a paired dataset of indoor low-light and normal-light images; S2. Perform feature extraction, filtering, and denoising on the collected remote sensing images; S3. Build an image enhancement model based on the improved Transformer network; S4. Design a loss function and set different training parameters to train the model to obtain an image enhancement model. S5. Input the low-light images in the test set into the image enhancement model for testing; S6. Obtain the enhanced image.

2. The low-light image enhancement method based on deep learning according to claim 1, characterized in that: The feature extraction module for feature extraction in step S2 adopts the RCBAM attention mechanism. The feature extraction module includes a channel attention module and a spatial attention module. Feature extraction includes a path that maintains the feature information of the original image, which is used to obtain high-level feature information when the image passes through the backbone network.

3. The low-light image enhancement method based on deep learning according to claim 1, characterized in that: The noise reduction is used to improve the purity of image processing and reduce noise interference.

4. The low-light image enhancement method based on deep learning according to claim 1, characterized in that: The loss function formula is as follows: ； The overall loss function formula is as follows: 。 5. The low-light image enhancement method based on deep learning according to claim 1, characterized in that: In step S2, a dual-path RCBAM attention mechanism is adopted, which includes a feature preservation path, an intelligent lighting adjustment module channel attention unit, and a spatial attention unit.

6. The low-light image enhancement method based on deep learning according to claim 5, characterized in that: The feature-preserving path establishes a direct connection channel, completely retains the original high-level features extracted by the backbone network, ensures that the basic information of the image is not lost through the residual connection, maintains the original brightness distribution characteristics, and prevents information attenuation.

7. The low-light image enhancement method based on deep learning according to claim 5, characterized in that: In the channel attention unit of the intelligent light adjustment module, a differentiated feature enhancement strategy is used to enhance the features of low-light areas by 3-5 times, and only slightly enhance the features of normal brightness areas by 1.2-1.5 times.

8. The low-light image enhancement method based on deep learning according to claim 5, characterized in that: In the spatial attention unit, a light adjustment heat map is generated through regional importance analysis, and the areas with enhanced intensity levels are dynamically divided.

9. A low-light image enhancement system based on deep learning, adopting the method described in any one of claims 1-8.

Images