Low-illumination image enhancement method suitable for complex night operation scene

Abstract

The invention discloses a low-illumination image enhancement method suitable for a complex night operation scene. The method comprises the following steps: preprocessing a collected low-illumination RGB image to obtain a preprocessed image; multi-scale features are extracted from the shallow convolution projection and the multilayer axial converter unit, and cross-layer attention is used for weighted aggregation to form enhanced features; forward noise addition is executed according to noise scheduling to obtain potential representation representing low-illumination noise distribution, and deep semantic representation is extracted in combination with a U-Net encoder embedded in CBAM; in the decoding stage, details are reconstructed through jump connection and a multi-layer converter, meanwhile, reverse denoising of a diffusion model is introduced to gradually remove noise, and an enhanced image with high brightness and low noise is output; in the training process, bidirectional mapping from low illumination to normal illumination is constructed, a weighted target of self-encoding loss and enhanced loss is introduced to carry out joint optimization on a converter and diffusion parameters, and a low-illumination scene is monitored on a general GPU / edge device.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, specifically to a low-light image enhancement method suitable for complex nighttime work scenarios. Background Technology

[0002] Low-light image enhancement is an important research direction in the field of computer vision. Its goal is to improve the visibility and quality of images in low-light environments, enabling them to be used more effectively in subsequent tasks. In scenarios such as nighttime surveillance, autonomous driving, security monitoring, and port operations, low-light images often suffer from insufficient brightness, color distortion, low contrast, and severe noise interference, making it difficult to distinguish image details and seriously affecting the performance of intelligent sensing systems.

[0003] Traditional low-light image enhancement methods mainly fall into two categories: those based on image processing and those based on biological visual mechanisms. The former, such as histogram equalization, enhances overall brightness by increasing contrast, but is prone to over-enhancement, loss of detail, and artifacts. The latter, such as retinal-cortical theory methods, decomposes the image into reflectance and illumination components to restore the image under normal lighting, but often leads to color shifts and noise amplification.

[0004] With the development of deep learning, convolutional neural network (CNN) methods have been widely applied to low-light image enhancement tasks. These methods achieve end-to-end learning to map low-light images to normal-light images, resulting in enhancement effects superior to traditional methods. However, due to the limited receptive field of CNNs, they are insufficient in capturing long-range dependencies and global illumination consistency, easily leading to insufficient detail recovery in the enhanced image. In recent years, the Transformer model, due to its advantages in global modeling, has been introduced into low-light image enhancement tasks, exhibiting better performance in image structure and detail recovery; however, its computational complexity is high, posing challenges for practical deployment.

[0005] In specific application scenarios, such as nighttime operations at container ports, low-light image enhancement faces even more complex challenges. Port yards have limited lighting coverage, with significant localized under-illumination, accompanied by strong noise interference, making it difficult to identify key targets such as vehicles, containers, and personnel. Furthermore, the port environment possesses complex visual features, such as ground reflections, water stains, stacked containers, and large machinery. These features are easily confused with background noise under low-light conditions, making them difficult to handle effectively with traditional methods.

[0006] In summary, existing low-light image enhancement techniques have the following shortcomings:

[0007] 1. Traditional image processing-based methods have limited enhancement effects and are prone to over-enhancement, color distortion, and noise amplification.

[0008] 2. Deep learning methods based on convolutional neural networks are limited by local receptive fields, making it difficult to capture global illumination changes and long-distance dependencies, resulting in insufficient detail recovery.

[0009] 3. Deep learning methods often rely on large-scale paired training data, but in practical engineering applications, high-quality paired data is difficult to obtain, which limits the generalization ability of the model.

[0010] 4. While attention-based or transformer-based models can model global dependencies, they have high computational overhead and are difficult to balance quality enhancement and operational efficiency in resource-constrained scenarios such as port monitoring.

[0011] In view of this, the present invention proposes a low-light image enhancement method suitable for complex nighttime operation scenarios. Summary of the Invention

[0012] The purpose of this invention is to provide a low-light image enhancement method suitable for complex nighttime operation scenarios, aiming to solve the problems of insufficient detail recovery, limited global illumination modeling capability, high dependence on paired training data, and limited computational efficiency of existing low-light image enhancement methods in complex scenarios.

[0013] In a first aspect, the present invention provides a low-light image enhancement method suitable for complex nighttime work scenarios, comprising the following steps:

[0014] S101: Image Acquisition and Preprocessing: Extract a set of normal light image samples and acquire low-light RGB images under low light conditions. Perform noise suppression, normalization and abnormal lighting correction on the low-light RGB images to form preprocessed images.

[0015] S102: Feature extraction and cross-layer fusion: The preprocessed image is input into a shallow convolutional projection layer to obtain initial features; it is then sequentially input into a multi-layer axis transformer unit to extract multi-scale features, and the features at each scale are weighted and aggregated through a cross-layer attention fusion layer to generate enhanced features;

[0016] S103: Forward Diffusion and Backbone Encoding: The enhanced features are input into the forward noise-adding unit of the diffusion model, and noise is injected step by step under a preset noise schedule to obtain noise modeling features; the noise modeling features are input into the encoder of the U-Net-type backbone structure, and deep features are extracted through downsampling and axis transformer units at each encoding stage, and convolutional attention modules (CBAM) are embedded in each module to enhance the representation ability of channel and spatial dimensions;

[0017] S104: Decoding Reconstruction and Backdiffusion: The encoded latent features are input into the decoder, and upsampling and multi-layer transformer operations are performed sequentially. The encoded features are then fused through skip connections. In the decoding stage, a backdiffusion denoising process is introduced to gradually remove noise and reconstruct clear image features. Finally, an enhanced high-brightness image is generated through a convolutional projection layer.

[0018] S105: Training and Consistency Constraints: During training, an enhancement mapping module from low-light images to normal-light images and an inverse mapping module from normal-light images to low-light images are constructed; an autoencoder loss function (ALF) is introduced to constrain the consistency between the input image and the enhancement-inverse mapping reconstruction result; the weighted sum of enhancement loss and autoencoder loss is used as the objective function to jointly optimize the transformer parameters and diffusion model parameters.

[0019] As a preferred embodiment of the present invention, the feature extraction and fusion step is specifically implemented in the following manner:

[0020] The transformer structure consists of multiple axial transformer units, and the preprocessed image is sequentially passed through multiple axial transformer units to extract multi-scale features.

[0021] The enhanced features are generated by weighted aggregation of multi-scale features through a cross-layer attention fusion layer.

[0022] As a preferred embodiment of the present invention, the forward diffusion process in step S103 is as follows:

[0023] The enhanced features are input into the forward noise-adding unit of the diffusion model, and random noise is gradually superimposed according to the preset noise schedule to simulate the noise distribution characteristics under low light conditions, thereby obtaining a potential representation containing noise perturbations at multiple time points; and the potential representation is used as the noise modeling feature.

[0024] As a preferred technical solution of the present invention, the deep feature extraction logic of step S103 is as follows:

[0025] The noise modeling features are input into the encoder with a U-Net backbone structure, and deep features are extracted through downsampling and axis transformer units at each encoding stage. Convolutional attention modules are embedded in each encoding module to enhance the representation ability of channel and spatial dimensions.

[0026] As a preferred technical solution of the present invention, the decoding and reverse diffusion logic of step S104 is as follows:

[0027] The encoded latent features are input into the decoder, and upsampling and multi-layer transformer operations are performed sequentially. The encoded features are then fused through skip connections. In the decoding stage, a reverse denoising process using a diffusion model is introduced to gradually remove noise and reconstruct clear image features. Finally, an enhanced high-brightness image is generated through a convolutional projection layer.

[0028] As a preferred embodiment of the present invention, the noise scheduling adopts a hybrid strategy of linear and cosine; during the training process, the step size allocation is switched according to the statistical results of the verification error to reduce the accumulation of back-diffusion error.

[0029] The bidirectional mapping module in step S105 includes constructing two generated mapping paths during the training process:

[0030] Forward mapping: Maps a low-light image to a normal-light image and calculates the difference in brightness distribution between the low-light image and the normal-light image samples;

[0031] Inverse mapping: Mapping a normal light image to a low light image and calculating the structural similarity difference between the two images and the input low light image;

[0032] Among them, the difference in brightness distribution is calculated by mean square error or histogram cross-entropy, and the difference in structural similarity is calculated by structural similarity index or perceptual loss function; the two differences are weighted and fused to form a bidirectional consistency constraint function.

[0033] As a preferred embodiment of the present invention, the parameters for joint optimization and synchronous adjustment include at least:

[0034] Noise scheduling and step size allocation in the diffusion model and noise estimator weights for inverse denoising;

[0035] Cross-layer attention weights, inter-layer fusion coefficients, and location coding parameters of the transformer;

[0036] The convolution kernel weights, downsampling and upsampling unit parameters, and gating coefficients of skip connections in the Net encoder and decoder;

[0037] The statistical update parameters of the normalized unit and the curve parameters of the hue / gamma mapping;

[0038] The weight coefficients of each component of the objective function include the brightness distribution term, the structural similarity term, the perception term, and the color constancy term.

[0039] As a preferred embodiment of the present invention, the objective function includes color constancy regularization and exposure consistency regularization: color constancy regularization is measured by the grayscale world deviation after channel normalization, and exposure consistency regularization is measured by the cross-entropy of the reference histogram; the weights of the two increase linearly in the early stages of training.

[0040] The two types of updates are performed alternately in odd and even training rounds, and the enhanced results are output when the convergence criterion is met.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] This invention effectively simulates the complex noise distribution in low-light environments by introducing a forward noise addition process using a diffusion model before the encoding layer, thereby improving the robustness of the model. Furthermore, by embedding spatial attention and channel attention mechanisms into the multi-layer transformer structure, it maintains global illumination consistency while enhancing local edge and texture details, overcoming the shortcomings of traditional self-attention methods that neglect local features. Simultaneously, a multi-step backward diffusion process is introduced in the decoding stage, enabling image recovery to gradually reduce noise from coarse to fine, avoiding over-smoothing and artifacts. By constructing forward and inverse mappings between low-light and normal-light images, and introducing bidirectional consistency constraints based on brightness distribution and structural similarity, the model can still converge stably even without paired training data. Therefore, this invention not only achieves superior performance compared to existing methods in objective indicators such as peak signal-to-noise ratio and structural similarity, but also outputs enhanced images with high brightness, low noise, and clear details, making it particularly suitable for complex low-light scenarios such as port nighttime operations and security monitoring. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a model architecture diagram of the present invention based on the fusion of diffusion model and converter;

[0045] Figure 2 This is a flowchart of the axial converter unit processing of the present invention;

[0046] Figure 3 This is a flowchart of the low-light image enhancement method of the present invention;

[0047] Figure 4 This is a schematic diagram of the bidirectional collaborative enhancement framework and cyclic consistency of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0049] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] Please see Figure 3 As shown, this embodiment provides a low-light image enhancement method suitable for complex nighttime operation scenarios. Addressing the complex nighttime environmental conditions and insufficient lighting in ports, it ensures that monitoring images remain clearly visible under low-light conditions, thereby improving operational safety and efficiency, and reducing monitoring blind spots and safety hazards caused by image blurring or obstruction. The method includes the following steps:

[0052] S101: Image acquisition and preprocessing: Extract a set of normal light image samples and acquire low-light RGB images under low light conditions. Perform noise suppression, normalization and abnormal lighting correction on the low-light RGB images to obtain preprocessed images.

[0053] It should be noted that: using low-light color images captured by port monitoring cameras or vehicle-mounted cameras as input, this study addresses the common problems of insufficient illumination, strong noise interference, and lack of texture details in low-light scenes. Based on the advantages of the diffusion model in modeling noise distribution, structural improvements were made to address the limitations of the transformer's dependence on large-scale paired data, thereby achieving efficient image enhancement under unpaired conditions.

[0054] The normal light image sample set consists of RGB images acquired under natural or uniform lighting conditions. These images cover port areas, cargo yards, vehicle lanes, and personnel work areas that are the same as or similar to low-light scenes, and are used as enhancement target domain samples during the training phase. The normal light image sample set provides a reference for cross-domain mapping for the network model, ensuring that the enhanced low-light images can approximate the brightness distribution and structural features of normal light images.

[0055] The processing logic for the preprocessed image is as follows:

[0056] Noise suppression: Adaptive noise filtering is performed on the acquired low-light RGB images, and spatial correlation and frequency domain analysis methods are used to remove Gaussian noise, stripe noise and impulse noise caused by insufficient illumination;

[0057] Normalization: The pixel intensity of the low-light RGB image after noise suppression is normalized, and the original pixel values ​​are linearly mapped to a fixed interval [0,1] or [−1,1] to reduce the difference in input distribution.

[0058] Abnormal illumination correction: Based on the local region histogram and luminance mask, compensation and adjustment are performed on areas that are too dark or too bright, and smoothing constraints are used to suppress boundary artifacts, thereby obtaining a preprocessed image with a balanced illumination distribution.

[0059] Understandably, after completing channel correction and noise suppression, brightness equalization can be performed on the entire image. Unlike traditional histogram equalization, this embodiment uses a zoned brightness adjustment method to improve global brightness while avoiding over-enhancing local areas, thus making the contrast of the pre-processed image more reasonable, highlighting dark information without causing overexposure of bright areas.

[0060] It should be understood that the preprocessed image obtained through the above processing will be passed as input to the subsequent shallow convolutional projection layer and axis transform unit. Since the image has been optimized in terms of color channel equalization, noise suppression effect and brightness distribution, it can more effectively express the real details in low-light scenes in subsequent feature extraction.

[0061] S102: Feature extraction and cross-layer fusion: The preprocessed image is input into a shallow convolutional projection layer to obtain initial features; it is then sequentially input into a multi-layer axis transformer unit to extract multi-scale features, and the features at each scale are weighted and aggregated through a cross-layer attention fusion layer to generate enhanced features;

[0062] It should be noted that even after structured preprocessing, low-light images still suffer from unclear details and weakened global structure. To ensure that the subsequent diffusion model can fully utilize effective features, this embodiment designs multi-scale feature extraction and cross-layer fusion steps. Local and global features of the image are extracted level by level using a combination of shallow convolutional projection layers and multi-layer axial transform units, and then weighted and aggregated through a cross-layer feature fusion unit.

[0063] To further explain, the preprocessed image is input into a shallow convolutional unit, such as... Figure 1 The 3×3 convolutional projection layer shown on the left yields the initial feature F0. The shallow convolutional units map the pixel spatial information of the preprocessed image to a high-dimensional feature space, facilitating effective processing by the subsequent attention mechanism. The initial feature F0 is sequentially input into multiple axis transformer units, each of which internally... Figure 2 As shown, it includes:

[0064] Height-axis multi-head attention: used to model global dependencies along the vertical direction, capturing vertical lighting changes and object contour relationships;

[0065] Width-axis multi-head attention: used to model global dependencies along the horizontal direction, supplementing structural information in the horizontal dimension;

[0066] Dual-gated feedforward network: used for nonlinear transformation and information channel adjustment to ensure that multi-scale semantics are taken into account when fusing features.

[0067] For example, these axial transformer units are stacked sequentially, generating F1, F2, and F3 from shallow to deep layers, respectively. Shallow features emphasize texture details, while deep features focus on the overall scene structure. A cross-layer attention fusion layer is added after multi-scale feature generation to weightedly combine features from different depths. This fusion unit first calculates the correlation weights between features at each scale, and then assigns different weight ratios to detail features and global features during the aggregation process, thus forming a balanced enhanced feature F4. This weighted fusion method avoids the information redundancy problem caused by simple stitching or superposition. When edge information in dark areas of the image needs to be enhanced, shallow features have a larger weight in the fusion; while when the overall illumination and structural relationships of the image need to be restored, the weight of deep features is correspondingly increased.

[0068] S103: Forward Diffusion and Backbone Encoding: The enhanced features are input into the forward noise-adding unit of the diffusion model, and noise is injected step by step under a preset noise schedule to obtain noise modeling features; the noise modeling features are input into the encoder of the U-Net-type backbone structure, and deep features are extracted through downsampling and axis transformer units at each encoding stage, and convolutional attention modules (CBAM) are embedded in each module to enhance the representation ability of channel and spatial dimensions;

[0069] Specifically, the noise modeling feature F4′ corresponds to the output of the forward modeling process of the diffusion model, and the logic for obtaining the noise modeling feature is as follows:

[0070] The noise distribution of low-light images in the training set is statistically analyzed, and its mean, variance, and spectral energy distribution features are extracted to form a noise distribution model.

[0071] Based on the noise distribution model, a noise scheduling table containing multiple stages is constructed, with each stage corresponding to noise components of different intensities and frequencies;

[0072] During the diffusion process, noise is injected into the enhancement features step by step according to the schedule, forming a progressively degraded feature sequence. In the last stage, the resulting feature is the noise modeling feature, which not only retains the structural information of the enhancement feature, but also contains noise perturbation components that match the low-light environment.

[0073] For example, in this embodiment, the noise injected at each step of the forward diffusion is not completely random, but is constrained by a noise distribution model so that the frequency components of the noise are consistent with the dark grain noise and sensor thermal noise common in low-light images. Therefore, the noise modeling features obtained are closer to the degradation characteristics of real low-light images.

[0074] It should be understood that after completing the forward diffusion modeling, the obtained noise modeling features are input into the backbone encoder based on the U-Net structure. The backbone encoder consists of multiple coding stages, each of which includes downsampling operations and axis transformer units. The downsampling operations are used to progressively reduce the feature map size and improve feature abstraction capabilities; the axis transformer units continue to capture long-range dependencies in the horizontal and vertical directions to ensure that global structural information is still preserved in the low-resolution feature space.

[0075] This invention improves upon the traditional encoder-decoder structure by injecting a noise-adding process before the encoding layer. This effectively simulates the multi-source noise distribution in low-light images and forces the network to model and be robust to real noise during the feature extraction stage. Before the encoding layer, the forward noise-adding unit performs a progressive noise-adding operation on the input preprocessed image based on a diffusion model to simulate the multi-source noise characteristics present in a low-light port environment. This allows the encoder to learn a more robust representation during feature extraction.

[0076] During feature extraction, the encoding layer incorporates a channel-spatial joint attention unit, which consists of a channel attention branch and a spatial attention branch. The channel attention branch dynamically adjusts the weights of each channel in the feature map by statistically analyzing the response intensity distribution of different channels, thereby highlighting illumination-related channels. The spatial attention branch, on the other hand, focuses on enhancing dark details and edge structures by analyzing the local spatial relationships of the feature map. The combination of these two approaches effectively strengthens the model's sensitivity to structural and detail information under low-light conditions.

[0077] It should be noted that by combining forward diffusion modeling with encoder enhancement, the noise characteristics under low-light conditions were simulated. Furthermore, during the encoding stage, the synergistic effect of the axial transformer and channel-space joint attention resulted in a more robust and discriminative deep feature representation. These features will directly support image reconstruction in the subsequent decoding and backdiffusion stages, ensuring that the enhanced image has high quality in both detail and global aspects.

[0078] S104: Decoding Reconstruction and Backdiffusion: The encoded latent features are input into the decoder, and upsampling and multi-layer transformer operations are performed sequentially. The encoded features are then fused through skip connections. In the decoding stage, a backdiffusion denoising process is introduced to gradually remove noise and reconstruct clear image features. Finally, an enhanced high-brightness image is generated through a convolutional projection layer.

[0079] It should be noted that although the deep latent features obtained after the forward diffusion modeling and encoder enhancement steps contain global structural and noise distribution information of the low-light image, they are still in a compressed representation state and cannot be directly used for high-quality image output. Therefore, a decoding reconstruction and reverse diffusion step is set up to gradually eliminate the noise introduced by diffusion modeling while restoring the image spatial resolution and recovering detailed features. After the decoding layer, the reverse denoising unit performs a progressive diffusion inverse process on the feature map of the decoded output. By gradually removing noise and restoring texture and color information, the enhanced image has higher naturalness and clarity.

[0080] Understandably, the decoder employs a progressive upsampling approach to gradually amplify the spatial resolution of latent features. At each decoding stage, the feature map undergoes convolution operations and a multi-layer transformer module. Convolution operations are used to recover local spatial details, while the transformer module maintains the continuous representation of long-range dependencies. Through this combined approach, the decoder can not only recover image texture but also maintain structural consistency at the global illumination level.

[0081] During the decoding process, a skip connection structure corresponding to the encoder is set up. Specifically, intermediate features generated in the encoding stage are fused with features of the corresponding level in the decoder through skip paths. This mechanism ensures that shallow detailed features can be directly passed to the decoding stage, thereby avoiding edge blurring and texture loss caused by multiple downsampling.

[0082] For example, while performing step-by-step decoding, this embodiment introduces a reverse denoising process of the diffusion model at each decoding stage. Specifically, based on the noise scheduling table used during forward modeling, the noise components in the feature map are gradually reduced in reverse order; in each denoising operation, the decoder uses the learned residual estimation results to correct the current features, thereby gradually removing random interference and restoring the true lighting and texture information.

[0083] The fusion of backdiffusion and skip connections has a synergistic effect: the shallow features provided by skip connections contain more edge and texture details of the original image, while the backdiffusion process can eliminate residual noise on this basis. The combination of the two can generate image features that perform well in terms of brightness balance and detail fidelity.

[0084] Understandably, after the decoding and back-diffusion processes are complete, the final reconstructed features are input to the convolutional projection layer. This layer maps the multi-channel high-dimensional features to the RGB image space, thereby outputting an enhanced high-light image. This output image outperforms the input low-light image in terms of brightness distribution, detail clarity, and color reproduction, meeting the needs of subsequent applications such as visual recognition, object detection, or image analysis. In the later training process, a normal light image under normal lighting conditions is used as the training target, and the normal light image is used as a reference to train the mapping relationship, enhancing the low-light image to a normal light image, thus enabling applications in complex nighttime work scenarios.

[0085] S105: Training and Consistency Constraints: During training, an enhancement mapping module from low-light images to normal-light images and an inverse mapping module from normal-light images to low-light images are constructed; an autoencoder loss function (ALF) is introduced to constrain the consistency between the input image and the enhancement-inverse mapping reconstruction result; the weighted sum of enhancement loss and autoencoder loss is used as the objective function to jointly optimize the transformer parameters and diffusion model parameters.

[0086] It should be noted that although the enhanced image can be generated through decoding reconstruction and backdiffusion steps, in order to ensure that the generated result is consistent with the normal illumination domain image in terms of brightness distribution, texture details and structure restoration, and to avoid mode collapse or insufficient enhancement during training, an autoencoded loss constraint is introduced to constrain the bidirectional enhancement strategy between the enhancement and degradation processes. This addresses the problem that traditional transformer-type low-light enhancement models generally rely on paired datasets for supervised learning. However, this invention, through the autoencoded loss constraint, enables the model to achieve stable cross-domain mapping even without pairing, thereby significantly improving its applicability and generalization ability in actual port monitoring scenarios.

[0087] Bidirectional enhancement strategies include forward mapping strategies and reverse mapping strategies, wherein:

[0088] The forward mapping strategy maps low-light images to normal-light images, thereby enhancing low-light images into normal-light images that are close to normal-light conditions.

[0089] The reverse mapping strategy maps normal light images back to low light images, simulating the reverse process from normal light images to low light images, and realizing bidirectional modeling of image structure and brightness information.

[0090] A stable mapping between low-light and normal-light images is achieved by constraining the consistency of the forward and backward mapping strategies using an autoencoder loss constraint function. To ensure the consistency of the bidirectional collaborative structure even under unpaired training conditions, an autoencoder loss constraint function is introduced. This function strengthens the collaborative consistency between the forward and backward mappings by calculating the distance between the enhanced and original images. Figure 4 As shown.

[0091] Specifically, the logic for obtaining the bidirectional consistency constraint is as follows:

[0092] Based on the progressive inverse denoising image generated from the aforementioned interactive features, two generation mapping paths are constructed: specifically:

[0093] Forward mapping strategy path: Map low-light images to normal-light images, and measure the difference in brightness distribution between the forward mapping strategy results and the corresponding normal-light image samples;

[0094] Inverse mapping strategy path: Mapping a normal light image to a low light image; measuring the structural similarity difference between the inverse mapping strategy result and the input low light image;

[0095] Among them, the difference in brightness distribution can be calculated by mean square error or histogram cross-entropy, and the difference in structural similarity can be calculated by structural similarity index or perceptual loss function based on feature mapping.

[0096] The aforementioned differences in brightness distribution and structural similarity are weighted and fused to form a two-way consistency constraint function.

[0097] During training, the converter parameters and diffusion scheduling parameters are updated by backpropagation using the bidirectional consistency constraint function, thereby ensuring that the mapping between low-light images and normal-light images maintains consistency and reversibility in both brightness and structure.

[0098] Example 2

[0099] The parts not detailed in this embodiment are as shown in Embodiment 1. Taking a port as an example, a self-built port dataset (hereinafter referred to as G) and low-light datasets V1 (hereinafter referred to as V1) and V2 (hereinafter referred to as V2) were established, and the diffusion transformer method was evaluated on public datasets. Ten representative low-light image enhancement methods were compared, including six unsupervised learning algorithms and four supervised learning algorithms; specifically:

[0100] Network Model 1 is a retinal network used to simulate the light-sensing mechanism of the human retina and enhance the overall brightness of low-light images. While the "supervised learning" method improves the perceived brightness of the image, it does not preserve sufficient detail, and noise remains significant.

[0101] Network Model 2 is a low-light transformer network used for global modeling based on the Transformer structure to improve illumination distribution. While it achieves good global consistency using a supervised learning method, it has limited ability to suppress complex noise and is prone to texture blurring.

[0102] Network Model 3 is a retinal transformer that combines the retinal perception mechanism with the Transformer module, balancing brightness enhancement and feature modeling. The "supervised learning" method improves both brightness and structural clarity, but increases network complexity and reduces inference efficiency.

[0103] Network Model 4 is a hue-luminance-intensity network, used to decompose the hue, luminance, and intensity components of an image, enhance them separately, and then fuse them. While employing a "supervised learning" method to achieve more natural color reproduction, it is prone to over-enhancement and local artifacts.

[0104] Network Model 5 is a Generative Adversarial Network (GAN) for lighting, used to learn the mapping relationship from low-light to normal-light conditions using GANs. While the "unsupervised learning" method enhances the realism of the results, adversarial training is unstable and lacks detail and texture recovery.

[0105] Network Model 6 is a zero-degree exposure correction network used to adjust the brightness of low-light images through an exposure correction mechanism. It adopts an "unsupervised learning" method to significantly improve brightness, but has limited detail recovery and insufficient generalization in complex port scenes.

[0106] Network Model 7 is a self-calibrating imaging network used to improve the quality of low-light images by utilizing a self-calibration mechanism. It employs an "unsupervised learning" method that has a certain effect on brightness improvement, but its noise suppression capability is weak and the target area is not clear enough.

[0107] Network Model 8 is a regressive uncertainty adaptive enhancement network used for adaptive enhancement methods based on uncertainty modeling. The "unsupervised learning" method improves brightness and contrast to a certain extent, but it is not good at suppressing background interference and the output image quality fluctuates greatly.

[0108] Network Model 9 is a global structure adaptive decoupling network used to introduce global structure constraints and decouple low-light images for modeling. It adopts an "unsupervised learning" method, which performs well in terms of global consistency, but lacks detail enhancement and has high computational complexity.

[0109] Network Model 10 is a paired low-light enhancement network, which is used for supervised training based on paired low-light and normal-light images. It adopts the "unsupervised learning" method and has a good enhancement effect under paired data conditions. However, due to the scarcity of paired data in practical applications, its generalization ability is limited.

[0110] This embodiment uses a diffusion transformer model, combining the diffusion model with the transformer structure. It introduces forward noise modeling noise before encoding and reverse diffusion denoising after decoding, and combines spatial-channel attention mechanism with autoencoding loss constraints.

[0111] Based on the above model, "S" and "U" represent "supervised learning" and "unsupervised learning" methods, respectively. The performance of the diffusion transformer method is comprehensively evaluated from two dimensions: computational efficiency and image quality. Four key evaluation indicators are used: peak signal-to-noise ratio (hereinafter referred to as SNR) and structural similarity index (hereinafter referred to as similarity ratio) are used to quantify image quality, while floating-point operations and parameter count measure the computational complexity of the model. The specific values ​​are shown in Table 1.

[0112] Table 1. Performance comparison of the present invention's model and existing models on port and low-light datasets:

[0113]

[0114] Experimental results show that in the port-based self-built dataset, the diffusion transformer model achieves a peak signal-to-noise ratio (PSNR) of 34.21 and a structural similarity ratio (SNR) of 0.948, significantly higher than other methods. Compared to the suboptimal method, the PSNR is improved by approximately 2.48%, and the SNR is improved by approximately 1.07%. In the low-light dataset V1, the diffusion transformer model achieves a PSNR of 25.56 and a SNR of 0.8497, outperforming other methods. Compared to the suboptimal method, the PSNR is improved by approximately 14.72%, and the SNR is improved by approximately 2.75%. In the low-light dataset V2, the diffusion transformer model achieves a PSNR of 29.17 and a SNR of 0.873, representing an improvement of approximately 4.11% in PSNR and approximately 0.58% in SNR compared to the suboptimal method.

[0115] The diffusion transformer model used in this embodiment outperforms the comparison methods in both PSNR and SSIM. Through diffusion-based inverse denoising, it preserves details of key areas such as vehicles, containers, and personnel in port scenarios. The autoencoder loss constraint ensures stable mapping even without paired training data, solving the problem of scarce paired data. Although it has more parameters than some lightweight models, it offers the best cost-effectiveness in terms of performance improvement. It is particularly suitable for complex low-light environments such as port nighttime monitoring, and can output high-brightness, low-noise, and detail-rich enhanced images in real time, improving operational safety and efficiency.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A low-light image enhancement method suitable for complex nighttime work scenarios, characterized in that, Includes the following steps: S101: Image acquisition and preprocessing: Noise suppression, normalization, and abnormal lighting correction are performed on low-light RGB images to form preprocessed images; S102: Feature extraction and cross-layer fusion: The preprocessed image is input into the transformer structure, and enhanced features are generated through multi-layer feature extraction and cross-layer fusion; S103: Forward Diffusion and Backbone Encoding: The enhanced features are input into the forward noise-adding unit of the diffusion model to obtain noise modeling features; The noise modeling features are input into the encoder with a U-Net-like backbone structure to extract deep features; S104: Decoding Reconstruction and Backdiffusion: The encoded latent features are input into the decoder, and a backdiffusion denoising process is introduced in the decoding stage to generate an enhanced high-brightness image. S105: Training and Consistency Constraints: A bidirectional mapping module is introduced to construct an autoencoder loss function to constrain and enhance consistency; the transformer parameters and diffusion model parameters are jointly optimized using the weighted sum of the enhancement loss and the autoencoder loss as the objective function.

2. The low-light image enhancement method for complex nighttime work scenarios according to claim 1, characterized in that, The feature extraction and fusion steps are implemented in the following ways: The transformer structure consists of multiple axial transformer units, and the preprocessed image is sequentially passed through multiple axial transformer units to extract multi-scale features. The enhanced features are generated by weighted aggregation of multi-scale features through a cross-layer attention fusion layer.

3. The low-light image enhancement method for complex nighttime work scenarios according to claim 2, characterized in that, The forward diffusion process in step S103 is as follows: The enhanced features are input into the forward noise-adding unit of the diffusion model, and random noise is gradually superimposed according to the preset noise schedule to simulate the noise distribution characteristics under low light conditions, thereby obtaining a potential representation containing noise perturbations at multiple time points; and the potential representation is used as the noise modeling feature.

4. The low-light image enhancement method for complex nighttime work scenarios according to claim 3, characterized in that, The deep feature extraction logic in step S103 is as follows: The noise modeling features are input into the encoder with a U-Net backbone structure, and deep features are extracted through downsampling and axis transformer units at each encoding stage. Convolutional attention modules are embedded in each encoding module to enhance the representation ability of channel and spatial dimensions.

5. The low-light image enhancement method for complex nighttime work scenarios according to claim 4, characterized in that, The decoding and reverse diffusion logic in step S104 is as follows: The encoded latent features are input into the decoder, and upsampling and multi-layer transformer operations are performed sequentially. The encoded features are then fused through skip connections. In the decoding stage, a reverse denoising process using a diffusion model is introduced to gradually remove noise and reconstruct clear image features; finally, an enhanced high-brightness image is generated through a convolutional projection layer.

6. The low-light image enhancement method for complex nighttime work scenarios according to claim 5, characterized in that, The noise scheduling adopts a hybrid strategy of linear and cosine; during the training process, the step size allocation is switched according to the statistical results of the verification error to reduce the accumulation of back diffusion error. The bidirectional mapping module in step S105 includes constructing two generated mapping paths during the training process: Forward mapping: Maps a low-light image to a normal-light image and calculates the difference in brightness distribution between the low-light image and the normal-light image samples; Inverse mapping: Mapping a normal light image to a low light image and calculating the structural similarity difference between the two images and the input low light image; Among them, the difference in brightness distribution is calculated by mean square error or histogram cross-entropy, and the difference in structural similarity is calculated by structural similarity index or perceptual loss function; the two differences are weighted and fused to form a bidirectional consistency constraint function.

7. The low-light image enhancement method for complex nighttime work scenarios according to claim 6, characterized in that, The parameters for the joint optimization and synchronous adjustment include at least: Noise scheduling and step size allocation in the diffusion model and noise estimator weights for inverse denoising; Cross-layer attention weights, inter-layer fusion coefficients, and location coding parameters of the transformer; The convolution kernel weights, downsampling and upsampling unit parameters, and gating coefficients of skip connections in the Net encoder and decoder; The statistical update parameters of the normalized unit and the curve parameters of the hue / gamma mapping; The weight coefficients of each component of the objective function include the brightness distribution term, the structural similarity term, the perception term, and the color constancy term.

8. The low-light image enhancement method for complex nighttime work scenarios according to claim 7, characterized in that, The objective function includes color constancy regularization and exposure consistency regularization: color constancy regularization is measured by the gray world deviation after channel normalization, and exposure consistency regularization is measured by the cross-entropy of the reference histogram. The weights of both increase linearly in the early stages of training.

Images