Multi-Scale Image Reconstruction via Feature Splicing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recovery methods are limited in addressing complex image damage scenarios, such as camera motion, object motion, and noise-induced invisibility, as they are primarily designed for specific types of blurring or noise.
Innovation Solution
An image processing method that involves acquiring a to-be-processed image, zooming it out to a smaller size, performing feature reconstruction using a first codec, zooming in the resulting feature map, splicing it with the original image, and then performing further feature reconstruction using a target codec to obtain a clear target image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a related image recovery method is used for specific types of blurring or noise, then the processing effectiveness for that specific problem is improved, but the adaptability to complex image damage scenarios (camera motion, object motion, noise-induced invisibility) deteriorates
Solution Approach 1:
The patent applies universality by designing an image recovery method that can handle multiple types of image damage simultaneously. The system processes images affected by camera motion, object motion, and noise-induced invisibility through a unified multi-scale feature reconstruction framework, making the method adaptable to various complex scenarios rather than being limited to specific damage types
Solution Approach 2:
The patent segments the image recovery process into multiple scales (coarse, intermediate, and fine scales) using different coding units. This segmentation allows each scale to address specific aspects of image damage, with coarse scales handling large-motion blurring and fine scales handling noise-induced invisibility, thereby improving overall adaptability to complex damage scenarios
2Manufacturing precision
If traditional image recovery methods are used, then the device complexity is reduced, but the manufacturing precision of image reconstruction deteriorates
Solution Approach 1:
The patent employs nested doll by organizing coding units into hierarchical levels where first coding units process coarse features and second coding units process fine features. The output of coarser scales feeds into finer scales, creating a nested structure that systematically reconstructs image details at increasing precision while managing complexity through organized modularity
Solution Approach 2:
The patent introduces a multi-scale dimension to the image recovery process, transforming the traditional single-scale reconstruction into a multi-dimensional approach. By processing images at coarse, intermediate, and fine scales separately and then fusing results, the system achieves superior reconstruction quality without proportionally increasing complexity, as each scale operates independently before integration
Data Source
AI summary
A method includes zooming out a to-be-processed image from a first size to a second size and determining a first feature map based on at least a feature reconstruction on the to-be-processed image of the second size by a first codec of a first set of coding units. The first set of coding units includes at least a pair of encoder and decoder. The method includes zooming in the first feature map to the first size, performing a splicing processing on the to-be-processed image of the first size and the first feature map of the first size to obtain a spliced feature map, and performing a feature reconstruction on the spliced feature map by a target codec, to obtain a target image. The target codec includes a target set of coding units, and a first subset of the target set corresponds to the first set of coding units.