Low-Light Image Enhancement Using Multi-Band Near-Infrared Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional low-light image improvement methods using the Retinex theory suffer from noise and color information loss when reconstructing low-light RGB images to normal light RGB images, necessitating additional analysis and improvement.
Innovation Solution
A low-light image improvement apparatus and method that utilizes a correlation between near-infrared multi-band images and low-light RGB images, employing a cross-attention transformer unit to fuse feature maps and reconstruct reflectance and illumination, incorporating structural components from the near-infrared images to enhance image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If conventional Retinex-based methods are used to improve low-light images, then brightness is improved, but noise and color information loss occur
Solution Approach 1:
The patent segments the low-light image improvement task into two distinct components: reflectance estimation (preserving color information) and illumination estimation (improving brightness). By using separate neural network branches for each component and processing them independently before combining, the method avoids the information loss that occurs when treating brightness enhancement as a single unified process.
Solution Approach 2:
The patent introduces an intermediary approach by using a multi-scale attention mechanism that operates as a mediator between the reflectance and illumination components. This attention mechanism selectively integrates information across different scales and components, preserving color fidelity while enhancing brightness without direct harmful interaction between the two processing streams.
2Illumination intensity
If conventional Retinex-based methods are used to improve low-light images, then brightness is improved, but noise is introduced
Solution Approach 1:
The patent applies preliminary action by first estimating the reflectance component before enhancing illumination. The reflectance map, which contains the structural and color information, is estimated first and then used as a guide for the subsequent illumination enhancement process. This sequential approach prevents noise introduction by establishing a stable color foundation before applying brightness adjustments.
Solution Approach 2:
The patent implements feedback mechanisms through its loss functions and attention mechanisms that continuously monitor and adjust the processing of reflectance and illumination components. The multi-scale attention mechanism provides feedback loops that ensure noise suppression while maintaining brightness improvement, allowing the system to self-correct and prevent harmful noise artifacts.
3Manufacturing precision
If detailed structural information is reconstructed in low-light images, then image quality improves, but processing complexity increases
Solution Approach 1:
The patent resolves the complexity issue by introducing another dimension - the multi-scale attention mechanism operates across multiple spatial scales and feature dimensions simultaneously. This allows the system to capture detailed structural information at different levels of abstraction without linearly increasing processing complexity, as the attention mechanisms efficiently aggregate information across scales in a hierarchical manner.
Data Source
AI summary
Disclosed are a low-light image improvement apparatus and method. The low-light image improvement apparatus includes: an image component decomposition network module that analyzes a light image, a low-light image, and a mid-light image to decompose reflectance and illumination, respectively, wherein the mid-light image is generated using the light image and the low-light image; and a component improvement network module configured to include a mid-teacher network model that extracts a first feature map with improved reflectance and illumination of the mid-light image using the reflectance and illumination of the light image and a student network module that distills the extracted first feature map and then extracts a second feature map for the reflectance and illumination of the low-light image based on the distilled first feature map and acquires an image with improved light by reflecting a structural component of a multi-band near-infrared image in the second feature map.