Multi-Resolution Image Segmentation With Mixed-Bit Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional electronic devices using deep learning face issues with increased quantization errors and deteriorated performance when bit depth is decreased, and prolonged processing times when image resolution is increased, making them unsuitable for devices with limited operational capabilities.
Innovation Solution
The method involves converting input images into multiple resolutions and processing them using separate deep learning engines with different bit depths to maintain processing time and improve performance, specifically employing a first deep learning engine with high bit depth for detailed regions and a second with lower bit depth for less detailed regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If bit depth is decreased in conventional deep learning image segmentation, then processing speed is improved, but quantization error increases and segmentation performance deteriorates
Solution Approach 1:
The patent divides the image into multiple resolution versions (first resolution and second resolution) and processes them separately through different deep learning engines. This segmentation approach allows different parts of the image to be processed with appropriate bit depths, maintaining precision for important regions while enabling faster processing overall.
Solution Approach 2:
The patent applies different bit depths to different regions of the image based on their importance. High bit depth is used for detailed regions requiring precise segmentation, while lower bit depth is used for less critical regions, optimizing the balance between processing speed and segmentation precision locally.
2Measurement precision
If resolution of input image is increased, then segmentation performance of detailed portion is improved, but processing time is increased
Solution Approach 1:
The patent segments the high-resolution image processing task into multiple parallel streams with different resolution levels. By processing the image at both high and low resolutions simultaneously through different deep learning engines, the system captures detailed information where needed while maintaining overall processing efficiency.
Solution Approach 2:
The patent applies partial high-resolution processing only to regions where detailed segmentation is critical, rather than processing the entire image at maximum resolution. This selective approach maintains detailed portion segmentation precision while reducing overall processing time through strategic use of computational resources.
Data Source
AI summary
A method for image segmentation based on deep learning in an electronic device includes: acquiring, by a processor of the electronic device, an input image, converting, by the processor, the input image into a first image having a first resolution and a second image having a second resolution, performing, by the processor, first image processing for the first image using a first deep learning engine, performing, by the processor, second image processing for the second image using a second deep learning engine different from the first deep learning engine, and providing, by the processor, region segmentation information of the input image based on first region segmentation information associated with the first image processing and second region segmentation information associated with the second image processing.