Adaptive Image Super-Resolution Through Similarity-Based Model Switching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution imaging techniques using machine learning struggle with low inference accuracy due to insufficient similarity between teacher data and inference target images, particularly in varying imaging conditions and locations, leading to reduced high frequency component inference.
Innovation Solution
An image processing apparatus that calculates the similarity between current and previous images, selects appropriate teacher data, and generates a learning model to infer high frequency components, ensuring high definition image generation with improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a learning model is trained with a wide variety of images to ensure uniform inference accuracy across all types, then adaptability is improved, but inference accuracy for specific similar images deteriorates
Solution Approach 1:
The system dynamically adapts the learning model by selecting and retraining with teacher data based on the similarity between the inference target image and available teacher data. When high similarity is detected, the system uses a specialized model trained on similar images; when similarity is low, it falls back to a general model trained on diverse images. This dynamic adaptation resolves the contradiction by adjusting the model behavior based on real-time conditions.
Solution Approach 2:
The system changes the training parameters (teacher data selection criteria) based on the similarity metric between the inference target and available teacher data. By adjusting which teacher data to use for model training based on image similarity parameters, the system optimizes inference accuracy for specific cases while maintaining the capability to handle diverse image types through the general model.
2Measurement precision
If teacher data with high similarity to the inference target is selected, then inference accuracy is improved, but adaptability to different imaging conditions deteriorates
Solution Approach 1:
The system dynamically switches between specialized and general learning models based on the similarity assessment. When the inference target image shows high similarity to available teacher data, it uses a specialized model for high accuracy. When imaging conditions vary significantly and similarity is low, it uses a general model trained on diverse images, thus adapting to different conditions while maintaining the option for high accuracy when applicable.
Solution Approach 2:
The system segments the learning model usage into two distinct paths: one for high-similarity cases using specialized teacher data, and another for low-similarity cases using general teacher data. This segmentation allows the system to optimize for accuracy when possible while maintaining adaptability for varying conditions through the general model path.
3Measurement precision
If learning is performed using only images from the same imaging location, then inference accuracy for that location is improved, but versatility across different locations deteriorates
Solution Approach 1:
The system dynamically adapts its learning approach based on the imaging location and similarity of the inference target. When the inference target is from a location with available similar teacher data, it uses location-specific training for high accuracy. When the location differs significantly or no similar data exists, it falls back to the general model trained on diverse locations, thus maintaining versatility across different imaging locations.
Data Source
AI summary
An image processing apparatus generates a learning model for making an image high definition using teacher data selected from among a plurality of teacher data which use an image included in the first image group as one of a pair of images. The image processing apparatus calculates a degree of similarity between a current image selected as a high definition target from a second image group and a previous image which is a high definition target previous to the current image, infers high frequency components of the current image using the generated learning model when the degree of similarity is equal to or less than a threshold, and infers high frequency components of a current image using a learning model used to make the previous image high definition when the degree of similarity is greater than the threshold.