Image Super-Resolution Using Precomputed Projection Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image super-resolution techniques face challenges in efficiently handling high magnification factors and are computationally costly, often requiring large training datasets and relying on accurate registration processes.
Innovation Solution
The method learns linear upscaling projection matrices from a training set of images, transforms patches to a standardized orientation to reduce memory usage and computational time, and applies these matrices for fast and efficient upscaling of images across any magnification factor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional super-resolution algorithms are used, then image quality can be improved, but processing time increases significantly
Solution Approach 1:
The patent pre-computes and stores projection matrices during an offline training phase using high-resolution training images. These pre-computed matrices are then directly applied during online super-resolution processing, eliminating the need for complex iterative optimization during runtime and achieving real-time performance while maintaining quality.
Solution Approach 2:
The patent divides the input low-resolution image into overlapping patches and processes each patch independently using the pre-computed projection matrices. This segmentation approach enables parallel processing and significantly reduces computation time compared to processing the entire image at once.
2Measurement precision
If large training datasets are used, then mapping accuracy between LR and HR manifolds improves, but computational cost increases
Solution Approach 1:
The patent performs the computationally intensive training process offline using large datasets to learn accurate projection matrices. Once trained, the model can perform super-resolution with minimal computational resources during online operation, effectively separating the high-cost learning phase from the low-cost application phase.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
Image super-resolution (SR) generally enhance the resolution of images. One of SR's main challenge is discovering mappings between low-resolution (LR) and high-resolution (HR) image patches. The invention learns patch upscaling projection matrices from a training set of images. Input images are divided into overlapping patches (410), which are normalized (420,430) and transformed (440) to a defined orientation. Different transformations can be recognized and dealt with by using a simple 2D-projection. The transformed patches are clustered, and cluster specific upscaling projection matrices and corresponding cluster centroids determined during training are applied to obtain upscaled patches. The upscaled patches are assembled to an upscaled image.