Image Search Device One-to-One Feature Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The Bag of Features (BoF) method for image search often retrieves incorrect results due to high appearance frequency of feature values, leading to images with different numbers of local features being incorrectly matched, such as the Turkish flag being mistaken for the US flag due to varying star marks.
Innovation Solution
An image search system that selects image feature values based on proximities to query feature values, ensuring each image feature value is unique and not reused across query feature values, and generates scores based on these selections to prioritize accurate matches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the BoF method uses tf-idf to statistically process appearance frequency of image feature values, then search speed is improved, but search accuracy deteriorates due to incorrect matches when appearance frequency is high
Solution Approach 1:
The patent segments the image matching process into two distinct stages: a rough filtering stage using tf-idf statistical processing to quickly eliminate obviously unrelated images, and a precise matching stage using one-to-one correspondence verification to accurately identify similar images. This segmentation allows the system to benefit from both the speed of statistical processing and the accuracy of precise matching.
Solution Approach 2:
The patent performs preliminary statistical processing using tf-idf to pre-filter the image database before conducting detailed one-to-one correspondence matching. This preliminary action reduces the search space and eliminates candidates that are unlikely to be matches, thereby improving overall search efficiency while maintaining accuracy through the subsequent precise matching stage.
2Adaptability or versatility
If multiple image feature values are selected for each query feature value based on appearance frequency, then more potential matches are captured, but one-to-one correspondence is lost leading to incorrect retrieval
Solution Approach 1:
The patent divides the feature value selection process into two phases: initial multi-value selection to ensure comprehensive match coverage, followed by one-to-one correspondence establishment to ensure matching reliability. This segmentation allows the system to first cast a wide net for potential matches and then refine the selections to ensure accurate correspondence.
Solution Approach 2:
The patent implements a feedback mechanism where the selection of image feature values is adjusted based on the establishment of one-to-one correspondence. When multiple image feature values are initially selected for a query feature value, the system provides feedback to refine the selection and establish unique correspondences, thereby maintaining both coverage and reliability.
3Productivity
If images with different numbers of local features are compared using statistical processing, then processing efficiency is maintained, but incorrect matches occur between images with varying feature counts
Solution Approach 1:
The patent segments the comparison process into statistical evaluation using tf-idf for efficiency and subsequent one-to-one correspondence verification for precision. This allows images with different numbers of local features to be initially evaluated efficiently through statistical metrics, followed by precise matching that accounts for varying feature counts by establishing unique correspondences between query and image feature values.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Provided is an image search device for suppressing a probability that an image to be searched may be selected when the number of local features included in the image to be searched is greatly different from the number of local features included in a query image. The image search device is configured to: obtain a plurality of query feature amounts each indicating a local feature of an image serving as a query; select image feature amounts respectively corresponding to the respective query feature amounts so that one image feature amount is selected relative to one of the plurality of query feature amounts for each of a plurality of images to be searched from image feature amounts respectively indicating local features of each of the images to be searched, which are stored in storage means in advance for each of the images; generate a score indicating a degree of similarity between each of the plurality of images to be searched and the image serving as the query based on the image feature amounts selected relative to the plurality of query feature amounts; and select at least one image from the plurality of images to be searched based on the score generated for the image.