Image Super-Resolution Using Texture-Fused Upsampling Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image super-resolution methods using interpolation algorithms result in image quality loss due to jagged or blurred images, as they rely solely on pixel position offset information without considering texture features.
Innovation Solution
Fusing pixel position offset information and texture features in a weight matrix to obtain a target weight matrix for upsampling, using a combination of convolutional and fully-connected neural networks to improve image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional interpolation algorithms are used for upsampling, then the upsampling process is simple and fast, but image quality deteriorates due to jagged or blurred images
Solution Approach 1:
The patent changes the parameters used for weight calculation from only pixel position offset information to a combination of pixel position offset information and texture features. This parameter expansion allows the weight matrix to capture both spatial relationships and local image characteristics, thereby improving image quality while maintaining the upsampling process efficiency through learned parameter transformations.
Solution Approach 2:
The patent combines multiple types of information (pixel position offset information and texture features) to create a composite weight matrix. This composite structure integrates spatial positioning data with local texture characteristics, enabling the upsampling process to simultaneously consider geometric relationships and image content, thus resolving the contradiction between simplicity and quality.
2Device complexity
If only pixel position offset information is used to calculate weight matrix, then the calculation process is simple, but image quality loss occurs due to lack of texture feature consideration
Solution Approach 1:
The patent adds another dimension to the weight calculation by incorporating texture features alongside pixel position offset information. This dimensional expansion transforms the weight matrix from a purely spatial-based structure to one that also encodes local image characteristics, enabling more accurate pixel value prediction during upsampling without excessively increasing computational complexity.
Solution Approach 2:
The patent performs preliminary extraction of texture features from the input image before the upsampling process. By pre-computing and storing texture feature representations, the system prepares this additional information in advance, so that during the actual upsampling operation, the weight matrix can be constructed using both position and texture data without adding significant real-time computational burden.
Data Source
Figure 1~2\
Figure 3~4a
Figure 4b~5
AI summary
Embodiments of this application disclose an image super-resolution method to convert a low-resolution image into a high-resolution image for display, and to be applied to various electronic devices with a display function. Based on a case in which pixel position offset information is obtained, a texture feature of a low-resolution image is further extracted, a weight matrix in which the pixel position offset information and the texture feature are fused is obtained, and a high-resolution image is obtained based on the weight matrix. The pixel position offset information and the texture feature are fused in the target weight matrix in this solution. Therefore, an image quality loss caused when interpolation processing is performed by relying on only the pixel position offset information can be avoided, and image quality of the high-resolution image can be improved.