Similar Image Retrieval Using Existence Probability for Human Perception
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing similar image retrieval devices face challenges in accurately retrieving images that align with human perception, as the reliability of similarity calculations is influenced by the accuracy of feature amounts and can result in estrangement from human appearance, especially when feature amounts are of multiple types.
Innovation Solution
A similar image retrieval device that calculates a first existence probability of a pattern within a retrieved image using a statistical method and then determines similarity based on this probability and a second existence probability of the pattern in instance images, using expressions such as coexistence and non-coexistence probabilities to ensure the retrieved image aligns with human perception.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If feature amounts of multiple types are used for similarity calculation, then the comprehensiveness of image analysis is improved, but the reliability of similarity calculation deteriorates due to varying calculation accuracy across different feature types
Solution Approach 1:
The patent transforms the similarity calculation from direct feature amount comparison to probability-based comparison. By converting feature amounts into existence probabilities through statistical methods, the system standardizes the reliability baseline across different feature types (pixel values, shapes, sizes), allowing comprehensive multi-type feature analysis while maintaining calculation reliability through probabilistic normalization.
2Device complexity
If feature amounts are used directly for similarity calculation, then the computational simplicity is improved, but the alignment with human perception deteriorates due to estrangement from human appearance
Solution Approach 1:
The patent introduces existence probability as an intermediary between feature amounts and similarity calculation. Instead of directly comparing raw feature amounts, the system first converts them into probabilities that reflect human perception of pattern existence. This intermediary layer bridges the gap between mechanical computation and human visual perception, maintaining computational tractability while improving perceptual alignment.
3Productivity
If traditional similarity calculation methods are used, then the processing speed is improved, but the accuracy of retrieving human-perceived similar images deteriorates
Solution Approach 1:
The patent performs preliminary conversion of feature amounts into existence probabilities before the actual similarity calculation. By pre-processing the feature data into a standardized probabilistic format, the system enables faster and more accurate similarity comparisons later. This preliminary action ensures that the subsequent retrieval process operates on perceptually-relevant data, improving both speed and accuracy of finding human-perceived similar images.
Data Source
AI summary
A feature amount calculation unit 61 calculates a feature amount corresponding to a pattern of a lesion by analyzing an inspection image. A probability calculation unit calculates a first existence probability which is a probability of the pattern of a lesion existing within the inspection image, using a calculation expression. The calculation expression is created in advance by a statistical method on the basis of a relationship between the feature amount and the presence or absence of the pattern of a lesion within an image for learning which is visually determined by a person. A similarity calculation unit calculates a similarity between the inspection image and a case image on the basis of the first existence probability and a second existence probability which is a probability of the pattern of a lesion existing within the case image which is calculated by the statistical method similarly to the first existence probability.