Image Search Device Proximity-Based Feature Vector Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing image search methods, such as the BoF method, face accuracy issues due to the reliance on appearance frequency of visual words, which deteriorates with increased cluster size and clustering accuracy, leading to errors in determining visual words.
Innovation Solution
An image search device and method that selects image feature vectors based on proximities between query feature vectors and stored image feature vectors, generating a score based on these proximities to improve the accuracy of similarity evaluation between query and search target images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the BoF method uses appearance frequency of visual words for score calculation, then the search process is simple and fast, but the search accuracy deteriorates due to cluster size and clustering accuracy issues
Solution Approach 1:
The invention changes the scoring parameter from appearance frequency to proximity-based scoring. Instead of using the frequency of visual words (which is affected by cluster size), the patent calculates scores based on the proximity between query feature vectors and image feature vectors, thereby improving search accuracy while maintaining computational efficiency
Solution Approach 2:
The invention substitutes the statistical mechanism (appearance frequency counting) with a geometric mechanism (proximity calculation in feature space). By replacing the BoF frequency-based scoring with Euclidean distance or other proximity metrics between feature vectors, the system achieves more accurate similarity measurement without sacrificing search speed
2Reliability
If the number of image feature amounts corresponding to visual word is increased, then the representation becomes more comprehensive, but the error in determining visual word increases
Solution Approach 1:
The invention extracts and uses only the most relevant feature vectors by selecting those with highest proximity to query feature vectors. Instead of using all feature amounts corresponding to a visual word (which introduces noise and errors), the patent selectively extracts the most significant features based on proximity measurement, thereby improving determination accuracy
Solution Approach 2:
The invention applies different quality standards to different feature vectors. Rather than treating all feature amounts uniformly, the patent evaluates each feature vector's proximity to query features and assigns different weights or selection priorities, thereby improving overall accuracy by focusing on high-quality local features
Data Source
AI summary
An image search device is configured to: obtain a plurality of query feature vectors each indicating a local feature of an image serving as a query; select a plurality of image feature vectors respectively corresponding to the plurality of query feature vectors based on proximities between the plurality of image feature vectors respectively indicating a plurality of local features of each of a plurality of search target images to be searched, and the plurality of query feature vectors; generate an image score of the search target image based on a total sum of score elements each corresponding to the proximities between the selected plurality of image feature vectors and the query feature vectors corresponding to the selected image feature vectors; and select at least one of the plurality of search target images based on the image score.