Low-Light Image Enhancement via Multi-Scale Context Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current denoising and enhancement techniques for low-light images fail to generalize well from synthetic to real images and do not adequately consider noise, leading to suboptimal results in real-world applications.
Innovation Solution
A method employing an encoder-decoder neural network with a U-net architecture and a multi-scale context aggregating block, which includes convolutional layers, pooling operations for downsampling, bilinear interpolation for upsampling, and dilation layers to aggregate local and global context information, effectively improving low-light image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If denoising techniques are applied to remove noises in low-light images, then image quality is improved, but current approaches fail to generalize well from synthetic data to real images
Solution Approach 1:
The patent changes the training parameters and data distribution by training the neural network on real low-light images rather than synthetic data, allowing the model to learn noise patterns and enhancement characteristics from actual camera captures. This parameter change in data source enables the model to generalize effectively to real-world images while maintaining high image quality.
Solution Approach 2:
The patent creates a neural network model that copies and learns from the characteristics of real low-light images during training, rather than relying on synthetic data. By copying the statistical properties, noise patterns, and visual characteristics from real images, the model can accurately enhance and denoise real-world images without the generalization gaps that occur with synthetic data training.
2Reliability
If enhancement techniques are applied to improve perceptual quality of digital images, then image appearance is improved, but current approaches do not take noise into consideration
Solution Approach 1:
The patent merges denoising and enhancement functions into a single unified neural network model. The model simultaneously performs noise removal and image enhancement in one processing pass, rather than applying separate denoising and enhancement steps. This combined approach ensures that enhancement operations are performed on denoised data, preventing the introduction of new artifacts while maintaining perceptual quality.
Solution Approach 2:
The patent converts the harmful noise present in low-light images into a useful training signal. By training the neural network on real low-light images with their inherent noise patterns, the model learns to distinguish between noise and valid image content. This allows the model to enhance images while preserving authentic visual characteristics and avoiding the introduction of artificial artifacts.
3Reliability
If extensive denoising and enhancement processing is applied to low-light images, then image quality is improved, but processing time and computational complexity increase
Solution Approach 1:
The patent replaces traditional mechanical image processing techniques with a neural network-based approach. Instead of using conventional algorithms that require multiple iterative processing steps, the neural network learns direct mappings from noisy input images to enhanced output images. This substitution dramatically reduces processing time while maintaining or improving image quality through learned feature representations.
Solution Approach 2:
The patent performs preliminary training of the neural network model offline using extensive real low-light image data. During training, the model pre-learns optimal enhancement parameters and noise patterns. This preliminary action allows the model to make accurate predictions and perform effective denoising and enhancement in real-time during actual image processing, without requiring complex computational operations at inference time.
Data Source
AI summary
In an embodiment, a method includes receiving a low-light digital image; generating, by at least one processor, a resulting digital image by processing the low-light digital image with an encoder-decoder neural network comprising a plurality of convolutional layers classified into a downsampling stage and an upsampling stage, and a multi-scale context aggregating block configured to aggregate multi-scale context information of the low-light digital image and employed between the downsampling stage and the upsampling stage; and outputting, by the at least one processor, the resulting digital image to an output device. In the upsampling stage, spatial resolution increases by using a bilinear interpolation operation performed before every few convolutional layers to speed up the inference time of the network.