Burst Image Restoration via Shift Estimation and Base Frame Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image restoration methods using deep learning-based neural networks face challenges in effectively handling burst images with degradations such as shift, noise, and blur, as they often require extensive training data and struggle to generalize well to unseen input patterns.
Innovation Solution
A processor-implemented method that utilizes a shift estimation model and a noise estimation model based on neural networks to determine shifted and noise data, selects a base image, and synthesizes image frames based on this data to perform image restoration, incorporating optical flow maps and feature maps to enhance image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep learning-based neural networks are used for image restoration, then image quality can be improved, but the network requires extensive training data and struggles to generalize well to unseen input patterns
Solution Approach 1:
The patent segments the image restoration task into multiple specialized networks, each trained on specific degradation types (shift, noise, blur). This allows each network to become an expert in its domain while the system as a whole handles diverse degradations through selective network invocation based on input analysis
Solution Approach 2:
The system dynamically changes operational parameters by selecting different restoration networks based on the detected degradation type and severity. Instead of using a single fixed network, the system adapts by choosing the most appropriate pre-trained network for the specific input characteristics, improving generalization without requiring retraining
2Measurement precision
If a single neural network is trained for specific image reconstruction, then it can generate accurate output for training patterns, but it lacks ability to handle unseen input patterns effectively
Solution Approach 1:
The patent creates a universal restoration system that encompasses multiple specialized networks, each with high accuracy for its specific function. The system achieves multi-functionality by maintaining a library of networks trained on different degradation types and selecting the appropriate one based on input analysis, thus combining specialized accuracy with universal applicability
Solution Approach 2:
The system performs preliminary analysis of the input image to identify degradation characteristics before selecting the appropriate restoration network. This preliminary action enables the system to prepare and invoke the most suitable pre-trained network in advance, ensuring high accuracy for unseen patterns by matching them with the most appropriate specialized network
Data Source
AI summary
A method and apparatus for image restoration using burst images are provided, where the method includes receiving a burst image set including image frames, determining shifted data representing a shift level of each of the image frames using a shift estimation model, selecting a base image from the image frames based on the shifted data, and performing image restoration by synthesizing at least a portion of the image frames based on the base image.