Image Upscaling via Patch Regression and Self-Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image upscaling methods face challenges in generating high-resolution images from low-resolution inputs due to the under-constrained nature of predicting millions of pixel values based on a small number of input pixels, often resulting in noise and artifacts, especially in images contaminated with noise or JPEG artifacts.
Innovation Solution
The method employs robust patch regression based on in-place self-similarity, which involves obtaining a low frequency band image intermediate, upsampling the input image, and estimating a result image by eliminating noise using a regression model that learns a first-order approximation of a nonlinear mapping function from low-resolution to high-resolution image patches, with adaptive processing to handle noise and texture patches effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image upscaling methods are used, then the process is simple and fast, but the output contains noise and artifacts and lacks accuracy
Solution Approach 1:
The patent divides the image into patches and processes them independently through the regression model, allowing complex processing to be applied locally without affecting the entire image. This segmentation enables accurate noise reduction and artifact removal while maintaining processing feasibility.
Solution Approach 2:
The patent introduces an intermediary regression model that learns the mapping between low-resolution and high-resolution patches. This model acts as a mediator that captures complex relationships and patterns, enabling accurate upscaling without directly complex processing operations.
2Reliability
If simple interpolation methods are used, then the processing is fast and simple, but the output quality is poor with visible noise and artifacts
Solution Approach 1:
The patent performs preliminary actions by pre-processing the image to separate frequency components and preparing patch data for the regression model. This preliminary preparation enables more efficient processing during the main upscaling operation, reducing overall processing time while improving quality.
Solution Approach 2:
The patent changes parameters by transforming the image into different frequency domains and adjusting the scale of patch processing. These parameter transformations enable the system to handle complex quality requirements more efficiently, improving output quality without proportionally increasing processing time.
3Object-affected harmful factors
If robust patch regression with in-place self-similarity is applied, then noise and artifacts are reduced and accuracy is improved, but the processing becomes more complex
Solution Approach 1:
The patent employs in-place self-similarity where the regression model uses the input image itself as part of the training data for generating high-resolution patches. This self-service approach eliminates the need for external training datasets and complex pre-processing, reducing algorithmic complexity while maintaining robust noise and artifact removal capabilities.
Solution Approach 2:
The patent creates copies of image patches at different resolutions and uses these copies within the regression model to learn mapping relationships. This copying mechanism simplifies the algorithm by avoiding complex mathematical operations while still achieving robust noise and artifact removal through pattern recognition.
Data Source
AI summary
Methods and systems for denoising and artifact removal in image upscaling are disclosed. In one embodiment, a low frequency band image intermediate is obtained from an input image. An upsampled image intermediate is obtained from the input image by upsampling. A result image is estimated, based at least in part on the upsampled image intermediate, the low frequency band image intermediate, and the input image. The input image is of a smaller scale than the result image. The estimating the result image further includes eliminating from the result image noise that is present in the input image.