ANN Image Upscaling for Corrupted Low-Resolution Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image upscaling methods using artificial neural networks (ANNs) struggle to produce high-quality results from low-resolution and corrupted images, as they lack training data that accounts for the quality degradation present in real-world scenarios.
Innovation Solution
A two-stage training process where a first ANN is trained using corrupted image data and supervised by a second ANN trained on higher-quality data, aligning features to improve upscaling performance on corrupted inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If an ANN is trained only on corrupted low-resolution image data, then it can process real-world compressed images, but the visual quality and accuracy of upscaled images deteriorates due to information loss
Solution Approach 1:
The patent merges the training processes of two ANNs by using the feature vectors from the second ANN (trained on high-quality data) as supervision signals for the first ANN (trained on corrupted data). This combination allows the first ANN to learn from both corrupted and clean image characteristics, resolving the contradiction between adapting to corrupted inputs and maintaining high output quality.
Solution Approach 2:
The feature vectors extracted from the second ANN serve as an intermediary representation that bridges the gap between corrupted training data and high-quality output requirements. These intermediate features guide the first ANN to produce high-quality upscaled images even when trained on corrupted data, acting as a mediator between the conflicting requirements.
2Loss of energy
If training data is compressed for storage and transmission, then bandwidth usage is reduced, but information loss occurs which affects the visual quality of resulting upscaled images
Solution Approach 1:
The patent implements a feedback mechanism where the second ANN, trained on high-quality uncompressed images, provides feature vector guidance to the first ANN. This feedback loop compensates for the information loss introduced by compression, allowing the system to maintain high output quality while accepting compressed low-resolution inputs, thus resolving the contradiction between bandwidth efficiency and information preservation.
3Manufacturing precision
If a single ANN is trained on high-quality uncorrupted image data, then upscaled image quality is improved, but the ANN fails to handle corrupted or compressed real-world images effectively
Solution Approach 1:
The patent segments the training task into two specialized ANNs: the first ANN specializes in processing corrupted low-resolution images, while the second ANN specializes in learning from high-quality data. By dividing the functionality and using the second ANN's features to supervise the first, the system achieves both high quality output and robustness to corrupted inputs, resolving the contradiction between quality and adaptability.
Data Source
AI summary
Image data representing one or more images at a first resolution is received at a first artificial neural network (ANN). The image data is processed using the first ANN to generate upscaled image data representing the one or more images at a second, higher resolution. The first ANN is trained to perform image upscaling and is trained using first training image data representing one or more training images at the first resolution, the first training image data being at a first level of quality. The first ANN is also trained using features of a second ANN, wherein the second ANN is trained to perform image upscaling and is trained using second training image data representing one or more training images at the first resolution, the second training image data being at a second level of quality, higher than the first level of quality.