Inference Model Suitability Evaluation Through Image Feature Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Noise removal using inference models generated by machine learning training is not effective for all images, particularly when the structure of the target image significantly differs from the training image, leading to inadequate noise removal or other performance issues.
Innovation Solution
A model usage evaluation method and system that compares the feature amounts of a target image with those of the training image to determine the appropriateness of using the inference model, including steps to acquire and compare feature vectors and calculate distances or angles to evaluate the model's suitability for the target image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If an inference model generated by machine learning training is used for noise removal, then noise removal effectiveness is improved for images with similar structure to training images, but noise removal effectiveness deteriorates for images with significantly different structure from training images
Solution Approach 1:
The patent performs preliminary evaluation by comparing feature amounts of the target image with training images before applying the inference model. This preliminary action determines whether the model is suitable for the given image, preventing ineffective noise removal on mismatched images and allowing selective application of the model only when appropriate.
2Productivity
If the inference model is applied to all images regardless of structure similarity, then processing speed is improved, but noise removal quality deteriorates
Solution Approach 1:
The system performs a quick feature amount comparison as a preliminary step before applying the computationally intensive inference model. This preliminary filtering enables rapid assessment of model suitability, allowing the system to quickly reject mismatched images without full processing while applying detailed processing only to compatible images, thus balancing speed and quality.
Solution Approach 2:
The patent applies partial processing (feature amount comparison) to all images, and only applies the full inference model processing to images that pass the preliminary evaluation. This partial action approach avoids the excessive computational cost of applying the full model to all images while ensuring quality processing for suitable images.
3Measurement precision
If feature amount comparison is performed to evaluate model usage appropriateness, then model usage accuracy is improved, but system complexity increases
Solution Approach 1:
The patent introduces feature amounts as an intermediary representation between the target image and the inference model evaluation. Instead of directly comparing complex image structures or applying the full inference model for evaluation, the system uses extracted feature amounts as a simplified mediator that captures essential characteristics, enabling accurate evaluation with reduced complexity.
Data Source
Figure 1
Figure 2(a)~2(c)
Figure 3
AI summary
Provided is a technique that enables appropriate use of an inference model generated by machine learning training. A model usage evaluation method for evaluating use of an inference model that is generated by machine learning training and inputs information based on an image includes: a target image feature amount acquisition step of acquiring a feature amount of a target image used as an input to the inference model (S01, S02); a training image feature amount acquisition step of acquiring a feature amount of a training image used for training for generating the inference model (S03, S04); and an evaluation step of comparing the feature amount of the target image acquired in the target image feature amount acquisition step with the feature amount of the training image acquired in the training image feature amount acquisition step to evaluate the use of the inference model for the target image (S05 to S08).