Image Super-Resolution via Neural Network Fusion and Sharpness Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution reconstruction methods are hindered by registration errors due to computational intensity and susceptibility to motion and image quality issues, leading to suboptimal image quality in reconstructed images.
Innovation Solution
An image super-resolution reconstruction method that involves obtaining continuous YUV images, extracting luma and chroma images, selecting a reference luma image based on sharpness, performing image registration to calculate offsets, using a trained neural network for fusion to achieve higher resolution luma images, and interpolating chroma images to synthesize high-resolution YUV images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image registration methods are used for super-resolution reconstruction, then the reconstruction process can be completed, but registration errors occur due to computational intensity and susceptibility to motion and image quality issues
Solution Approach 1:
The patent introduces a deep learning model as an intermediary between image registration and fusion reconstruction. The model learns the mapping relationship from low-resolution images to high-resolution images, effectively bridging the gap and avoiding traditional registration errors. The learned transformation is then applied to guide the fusion process, achieving both accuracy and robustness.
Solution Approach 2:
The patent replaces the mechanical image registration system with a data-driven deep learning approach. Instead of relying on feature matching and geometric transformation calculations that are sensitive to motion and quality variations, the system uses neural networks to directly learn the super-resolution mapping, substituting computational mechanics with intelligent pattern recognition.
2Manufacturing precision
If image registration is performed to align multiple images for super-resolution, then higher resolution information can be utilized, but computational complexity increases significantly
Solution Approach 1:
The patent extracts and processes only the essential information needed for super-resolution. The deep learning model is trained to directly map low-resolution images to high-resolution images, extracting only the necessary high-frequency details and structural information rather than performing exhaustive registration and fusion of all image components.
Solution Approach 2:
The patent changes the fundamental parameters of the super-resolution approach by using learned transformation fields instead of traditional registration parameters. The model learns optimal transformation parameters during training that adapt to different image content and motion patterns, enabling high-resolution reconstruction with reduced computational burden during inference.
3Manufacturing precision
If traditional fusion reconstruction methods are used after registration, then the process can be completed, but registration errors are converted into reconstruction errors, adversely affecting image quality
Solution Approach 1:
The patent converts the potential harm of registration errors into benefit by using the deep learning model to learn from various motion patterns and image quality variations during training. The model is exposed to registered images with different alignment qualities and learns to produce high-quality outputs regardless of registration imperfections, effectively turning the registration process from a source of errors into a source of diverse training data that improves robustness.
Data Source
AI summary
The disclosure provides an image super-resolution reconstruction method, a mobile terminal, and a computer-readable storage medium. The method includes: obtaining continuous N first YUV images; extracting N luma images and N chroma images from the N first YUV images; performing sequentially sharpness estimation, image registration, and image reconstruction based on a neural network on the N luma images; performing image reconstruction on the N chroma images; and fusing the chroma image obtained after reconstruction and the luma image obtained after reconstructions to obtain the target YUV image that has a higher resolution than the N first YUV images.