Hybrid Low Light Image Enhancement Method
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mobile electronic devices face challenges in capturing high-quality images in low light environments due to limited processing power and energy resources, leading to unsatisfactory image quality or prolonged processing times that disrupt device performance.
Innovation Solution
A hybrid image enhancement method combining a low light enhancement model of a deep neural network with a sharpness-preserving mathematical model, which downsamples high-resolution images for efficient processing and upsamples them to maintain real-time image presentation while optimizing processor loading.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep neural network processing is applied to enhance low light images, then image quality is improved, but processing time increases
Solution Approach 1:
The image processing task is divided into two segments: a deep neural network component that handles low light enhancement on downsampled images, and a mathematical model component that performs sharpness-preserving upsampling on high resolution images. This segmentation allows each component to operate on appropriately scaled data, improving overall processing efficiency while maintaining image quality.
Solution Approach 2:
A downsampling operation serves as an intermediary step between the high resolution low light image and the deep neural network processing. By downsampling to a lower resolution before neural network processing, the computational load is reduced while still capturing essential semantic information, thereby decreasing processing time without significantly compromising quality.
2Measurement precision
If high resolution images are processed through deep learning models, then semantic detail recovery is improved, but processor loading increases
Solution Approach 1:
The patent extracts and removes the high resolution component from the deep learning processing pipeline. Instead of processing the full high resolution image through the deep neural network, only the downsampled lower resolution image is processed, extracting only the necessary semantic information at a reduced computational cost.
Solution Approach 2:
The system applies partial processing by using the deep neural network only on the downsampled image portion rather than the complete high resolution image. This partial action approach processes only the essential semantic content while avoiding the excessive computational load of processing every pixel at full resolution.
3Productivity
If image enhancement processing is performed in real-time, then device performance is maintained, but processing power consumption increases
Solution Approach 1:
The processing pipeline is segmented into a neural network component operating on downsampled images and a mathematical model component handling upsampling. This segmentation enables real-time processing by distributing computational tasks appropriately, reducing overall power consumption while maintaining productivity.
Solution Approach 2:
The system changes the resolution parameter by downsampling the input image before neural network processing. This parameter change reduces the number of pixels processed, thereby lowering processing power consumption while still maintaining the ability to produce real-time enhanced output through the subsequent upsampling stage.
Data Source
AI summary
A hybrid method of low light image enhancement in an electronic device comprises downsampling a high resolution, low light image to obtain a low resolution, low light image and inputting the low resolution, low light image into a low light enhancement model of a deep neural network to obtain a low resolution, enhanced image where semantic information such as true colors, edges, and brightness level are recovered. The high resolution, low light image, the low resolution, low light image, and the low resolution, enhanced image are each input into a mathematical model that performs upsampling and super-resolution tasks to generate a high resolution, enhanced image.