Image Retrieval Feature Descriptor Filtering via Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional content-based image retrieval (CBIR) systems face high computational complexity due to the need to compare large numbers of feature descriptors, making them slow and inefficient, especially when using descriptors like SIFT that generate thousands of features per image.
Innovation Solution
The method involves calculating a single global feature vector for each image, clustering images based on distance measures, filtering feature vectors that are matched across a threshold number of images, and using these filtered vectors to efficiently retrieve similar images, thereby reducing computational load and noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all feature descriptors from one image are compared with all feature descriptors from all other images, then comprehensive image similarity search is achieved, but computational complexity becomes extremely high
Solution Approach 1:
The patent segments the image database into multiple clusters based on global feature vectors. Instead of comparing all descriptors across the entire database, the system divides the search space into manageable clusters, reducing the number of comparisons needed while maintaining comprehensive search capability through cluster-based organization.
Solution Approach 2:
The patent performs preliminary clustering of images into clusters using global feature vectors before the actual similarity search. This preliminary organization allows the system to quickly narrow down candidate images to relevant clusters, avoiding unnecessary comparisons with images from unrelated clusters and significantly reducing computational complexity.
2Measurement precision
If SIFT descriptors are used to generate detailed feature vectors, then accurate local feature matching is achieved, but the number of descriptors per image increases to over 20000
Solution Approach 1:
The patent extracts and filters only the most relevant local feature descriptors from each image based on matching frequency across clusters. Instead of using all SIFT descriptors, the system identifies and retains only those descriptors that consistently match across multiple images within the same cluster, significantly reducing the number of descriptors while preserving matching accuracy.
Solution Approach 2:
The patent applies different processing strategies to different parts of the feature descriptor set. Global feature vectors are used for initial clustering, while local SIFT descriptors are selectively filtered and retained based on their matching frequency. This differentiated approach optimizes both computational efficiency and matching precision for different descriptor types.
3Reliability
If feature vectors from all images are stored and compared, then complete search coverage is maintained, but memory usage and processing time increase significantly
Solution Approach 1:
The patent performs preliminary filtering of feature vectors during the clustering phase, identifying and storing only those descriptors that exhibit high matching frequency within clusters. This preliminary selection creates a condensed representation of the image database that maintains search coverage for relevant features while dramatically reducing the data volume for subsequent search operations.
Solution Approach 2:
The patent segments the feature descriptor storage into cluster-specific filtered descriptors rather than maintaining a single comprehensive database of all descriptors. Each cluster stores only the filtered descriptors relevant to that cluster's images, enabling efficient localized searches without the overhead of processing the entire database for every query.
Data Source
AI summary
The present disclosure generally relates to methods and computer program products for searching for a similar image among a plurality of stored images, and in particular to a method and computer program product used in a content based image retrieval system where roughly similar images are clustered and feature vectors for the clustered images are filtered based on a matching frequency for the feature vectors among the images in the cluster.