A low-light image enhancement method based on a double-layer asymmetric network

Abstract

This invention discloses a low-light image enhancement method based on a two-layer asymmetric network, belonging to the field of computer vision and image processing technology. Addressing the problem of significant differences in the physical properties of intensity and chromaticity components in low-light images within the HVI color space, this invention proposes a two-layer asymmetric network architecture. First, the input image is decomposed into intensity and chromaticity components. Then, a dual-path parallel processing stream is constructed. The intensity stream uses a shallow network to retain high-frequency structural information, while the chromaticity stream uses a deep network to expand the receptive field and denoise. A cross-guided interaction module (CGIB) is used to asymmetrically process the two feature streams. A spectral decoupling strategy is introduced in the intensity branch to adjust the illumination, and multi-scale dilated convolution is introduced in the chromaticity branch to extract color context. Finally, bidirectional guided affine modulation (GAM) is used to fuse features and reconstruct the image. This invention effectively solves the problems of color shift and noise residue in low-light enhancement, while reducing computational redundancy and improving image enhancement results.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a low-light image enhancement method based on a two-layer asymmetric network. Background Technology

[0002] Low-light image enhancement aims to recover high-fidelity image signals from photon-scarce, noisy, and color-distorted inputs. Traditional deep learning methods typically operate in the sRGB domain, where the brightness and chromaticity of an image are non-linearly coupled. This entanglement forces a single network to simultaneously handle global illumination adjustment, local denoising, and color correction, often leading to optimization conflicts and resulting in artifacts such as color shifts or residual noise.

[0003] To address the coupling problem, existing techniques attempt to physically separate the image into intensity components using hypersphere (HVI) decomposition. ) and chromaticity components ( However, existing state-of-the-art (SOTA) methods typically employ a "symmetric paradigm," which treats intensity and chromaticity components with the same network depth and homogenized operators.

[0004] In reality, intensity components are typically dominated by low-frequency signals, corresponding to continuous global illumination, and require the preservation of clear structural edges. Chromatic components, on the other hand, exhibit a dispersed distribution, are heavily influenced by noise and color texture, and belong to high-frequency signals, requiring denoising using a wide range of contextual information. Existing symmetrical processing methods ignore this difference in physical characteristics: uniform deep networks may lead to the loss of structural information in intensity components, while uniform shallow networks cannot provide sufficient receptive field for denoising chromatic components. Therefore, how to design differentiated enhancement networks for the physical characteristics of different components is a pressing problem that current technology needs to solve. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides a low-light image enhancement method (DA-Net) based on a two-layer asymmetric network. By introducing asymmetric design at both the network topology and internal processing operator levels, it achieves efficient and accurate recovery of intensity and chromaticity components.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A low-light image enhancement method based on a two-layer asymmetric network includes the following steps:

[0008] Step S1: Project the input image into the HVI color space and separate the intensity component and chromaticity component.

[0009] Step S2: Construct an asymmetric dual-stream network architecture. The intensity stream is designed as a shallow network (e.g., downsampling only at 1 / 2 scale) as a structural anchor to preserve high-frequency spatial details; the chroma stream is designed as a deep U-Net architecture (e.g., downsampling up to 1 / 8 scale) to significantly expand the receptive field and utilize long-range context to distinguish true color signals from noise.

[0010] Step S3: Feature processing is performed using the cross-guided interaction module. This module internally implements asymmetric operations.

[0011] (1) Brightness refinement branch: For the intensity component, a strategy of depthwise convolution before pointwise convolution is adopted, and a spectral decoupling strategy is introduced. Through fast Fourier transform and learnable Gaussian soft mask, the features are separated into low frequency (illumination) and high frequency (structure) in the frequency domain. Illumination adjustment is performed only on the low frequency component, thereby avoiding structural blurring and ringing effect.

[0012] (2) Chromaticity refinement branch: For chromaticity components, an inverse residual strategy of pointwise convolution before depthwise convolution is adopted, and multi-scale dilated convolution is used to capture spatially diffused color information, combined with a gating mechanism to suppress noise.

[0013] Step S4: Fuse features using bidirectional guided affine modulation (GAM). High signal-to-noise ratio intensity features constrain the diffusion of chrominance features (preventing color overflow), while semantic information of chrominance features assists in identifying dark area structures using intensity features. Finally, the fused features are mapped back to the sRGB domain for output.

[0014] This invention breaks through the limitations of traditional symmetric processing paradigms, significantly improving image restoration quality and effectively eliminating color shifts and residual noise commonly found in low-light enhancement. It outperforms existing state-of-the-art methods in objective metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Secondly, this invention significantly reduces computational redundancy while improving performance. Because the intensity flow is designed as a shallow structure, unnecessary deep convolution calculations on high-resolution structural features are avoided, thereby greatly reducing the number of floating-point operations (FLOPs) in the model while maintaining high image fidelity, achieving an excellent balance between performance and efficiency. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0016] Figure 2 This is a schematic diagram of the Cross-Guided Interaction Module (CGIB) in this invention;

[0017] Figure 3 This is a schematic diagram of the bidirectional guided affine modulation (GAM) structure in this invention;

[0018] Figure 4 This is a schematic diagram of the attention gating unit (AGM) in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The low-light image enhancement method based on a two-layer asymmetric network proposed in this invention, such as... Figure 1 As shown, the low-light input image to be processed is first obtained.

[0021] The acquired low-light image is projected from the standard RGB color space to the HVI color space, and the intensity component is obtained by decoupling. and chromaticity components .

[0022] The signal characteristics of low-light images are manifested in the intensity component. The chromaticity component is dominated by low-frequency signals and corresponds to continuous global illumination. It is dominated by high-frequency noise and is dispersed.

[0023] For intensity components A shallow intensity flow processing path is constructed, which contains only a single downsampling operation.

[0024] The intensity stream processing path maintains the feature map resolution at half the scale of the input image through a single downsampling.

[0025] This shallow topology serves as the structural anchor point for the entire network, which can minimize the loss of sparse edge textures and high-frequency spatial details during feature extraction.

[0026] For chromaticity components A deep chroma stream processing path is constructed in parallel with the intensity stream, which adopts a multi-cascaded U-Net architecture.

[0027] The chroma stream processing path compresses the feature map resolution to one-eighth the scale of the input image through three consecutive downsampling operations.

[0028] During the decoding stage of the chroma stream, the feature map resolution is restored to the corresponding scale through step-by-step upsampling.

[0029] Deep topology enables networks to obtain a large effective receptive field, utilizing long-range semantic context information to accurately distinguish real color signals from random noise in noise-dense chroma components.

[0030] To enable information exchange between two heterogeneous flow paths, a cross-guided interaction module (CGIB) is designed, which includes a shared attention gating unit (AGM).

[0031] like Figure 4 AGM combines SimpleGate and Simple Channel Attention (SCA) for dynamically recalibrating features. The specific implementation formula for AGM is:

[0032]

[0033] The information channel is selectively emphasized through the global context.

[0034] like Figure 2 The Cross-Guided Interaction Module (CGIB) contains parallel luma refinement branches and chroma refinement branches.

[0035] Luminance refinement branch for intensity component It employs a connection strategy that prioritizes depthwise convolution over pointwise convolution to preserve spatial structure.

[0036] In the brightness refinement branch, a spectrum decoupling strategy is introduced to handle global illumination. First, the features are subjected to a fast Fourier transform.

[0037] Using learnable Gaussian soft masks The amplitude spectrum is separated into low-frequency illumination components and high-frequency structure components, using a Gaussian soft mask. Based on frequency coordinates The Euclidean distance to the center of the spectrum is calculated using the following formula:

[0038]

[0039] Learnable parameters for controlling spectral bandwidth. A constant used to ensure numerical stability.

[0040] After the separated low-frequency illumination component and high-frequency structural component are processed by independent convolution, they are restored to the spatial domain by inverse fast Fourier transform.

[0041] The chromaticity refinement branch targets chromaticity components. It employs an inverse residual connection strategy that uses pointwise convolution before depthwise convolution to decouple color representation.

[0042] The chroma refinement branch utilizes multi-scale dilated convolution to extract features, capturing spatially diffused color information and obtaining aggregated features. .

[0043] Combining gating mechanisms for aggregated features Weighting is performed using the formula. Automatically suppress noise-dominant channel responses, where This is the Sigmoid activation function.

[0044] like Figure 3 As shown, in the feature fusion stage, a bidirectional guided affine modulation (GAM) mechanism is introduced to unify the feature distribution. This module adopts a residual affine transformation mechanism.

[0045] set up To guide the mode, To achieve complementary target modalities, a multilayer perceptron is used to guide the features. Mapped to pixel-level modulation parameters, including scaling map Peace map .

[0046] Through formula , for target features Modulation is performed, in which This indicates element-wise multiplication. This ensures an identity mapping at zero initialization, guaranteeing training stability.

[0047] When high signal-to-noise ratio intensity flow features are used as guides, the chrominance flow features are modulated by modulation formulas. The clear boundaries of intensity are used to constrain the spatial diffusion of chrominance, preventing color overflow at the edges of the enhanced image.

[0048] When chroma flow features are used as a guide, the coherent color texture semantics contained therein are fed back to the intensity flow, and the modulation formula helps the intensity flow distinguish noise from real texture in extremely low illumination areas.

[0049] Features that have undergone asymmetric processing and interaction are fused and mapped back to RGB space through an inverse projection module to obtain the final high-quality enhanced image.

Claims

1. A low-light image enhancement method based on a two-layer asymmetric network, characterized in that, Includes the following steps: Step S1: Obtain the low-light image to be processed, project it from the RGB color space to the HVI color space, and decouple it to obtain the intensity component and chromaticity component; Step S2: Input the intensity component and chromaticity component into the parallel intensity stream processing path and chromaticity stream processing path in the two-layer asymmetric network, respectively; The intensity stream processing path employs a shallow network structure to preserve the spatial structural features of the intensity components; the chroma stream processing path employs a deep encoder-decoder structure with a greater number of network layers than the intensity stream processing path, in order to expand the receptive field and remove noise from the chroma components. Step S3: During the feature extraction process of the intensity stream processing path and the chroma stream processing path, the cross-guided interaction module is used to perform differentiated extraction and interaction of the two features; Step S4: Use the bidirectional guided affine modulation module to fuse the processed intensity and chromaticity features, and then back-project the fused features back into the RGB color space to obtain the enhanced image.

2. The method according to claim 1, characterized in that, In step S2, the topology of the two-layer asymmetric network is specifically as follows: The intensity stream processing path contains only a single downsampling operation, maintaining the feature map resolution at half the scale of the input image; The chroma stream processing path includes multiple cascaded downsampling operations, which progressively compress the feature map resolution to 1 / 8 of the input image scale, and progressively upsample and restore it during the decoding stage.

3. The method according to claim 1, characterized in that, The Cross-Guided Interaction Module (CGIB) includes a shared Attention Gating Unit (AGM) for dynamically recalibrating features; the AGM combines simple gating with simple channel attention, selectively emphasizing information channels through global context.

4. The method according to claim 3, characterized in that, The cross-guided interaction module includes parallel luminance refinement branches and chrominance refinement branches: The brightness refinement branch: For the intensity component, a connection strategy of depthwise convolution before pointwise convolution is adopted to preserve the spatial structure; then, a spectral decoupling strategy is used to process global illumination; The chroma refinement branch employs an inverse residual connection strategy, prioritizing pointwise convolution over depthwise convolution, to decouple color representation; and utilizes multi-scale dilated convolution to extract features.

5. The method according to claim 4, characterized in that, The spectrum decoupling strategy specifically includes: Perform a Fast Fourier Transform on the features and utilize a learnable Gaussian soft mask. The amplitude spectrum is separated into a low-frequency illumination component and a high-frequency structure component; the Gaussian soft mask Based on frequency coordinates Euclidean distance to the center of the spectrum: in, Learnable parameters for controlling spectral bandwidth. The constant is used; after the separated components are processed by independent convolution, they are restored to the spatial domain by inverse fast Fourier transform (IFFT).

6. The method according to claim 1, characterized in that, The bidirectional guided affine modulation module (GAM) employs a residual affine transformation mechanism: set up To guide the mode, For complementary target modalities; using a multilayer perceptron to guide features Mapped to pixel-level modulation parameters, including scaling map Peace map ; The target features are determined using the following formula. Modulation: in This indicates element-wise multiplication. This ensures an identity mapping at zero initialization.

Images