Image Ranking for Visual Search Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing visual search methods in product databases yield inaccurate results due to the inclusion of irrelevant or misleading images, particularly those showing product applications or additional objects, leading to false positives.
Innovation Solution
A method is introduced to generate a ranked database by determining a reference image for each product data set, calculating similarity measures, and assigning image ranks based on these measures to prioritize relevant images, using techniques like deep learning and cosine similarity, thereby excluding less relevant images from search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all images associated with a product are included in the database for visual search, then the search database is comprehensive, but false positive results increase due to irrelevant images
Solution Approach 1:
The patent segments the image database by assigning hierarchical ranks to images (first rank, second rank, etc.) based on their relevance to the product. This segmentation allows the search system to prioritize high-rank images while still maintaining access to lower-rank images, thus improving search accuracy without discarding any images from the database.
Solution Approach 2:
The patent extracts and identifies the most relevant images (first rank images) from the complete set of product images. By separating these high-value images into a distinct category, the system can focus search matching on these extracted images, thereby reducing false positives while maintaining database completeness.
2Loss of information
If multiple images per product are stored in the database, then product representation is comprehensive, but search processing complexity increases
Solution Approach 1:
The patent performs preliminary ranking of all images during database ingestion, assigning each image a rank based on its relevance to the product before search operations occur. This preliminary action organizes the images in advance, so that during search, the system only needs to compare the query image against images in ascending rank order, significantly reducing processing complexity compared to evaluating all images equally.
Solution Approach 2:
The patent implements a dynamic search process where the system automatically stops comparing images once it reaches a certain number of matches or exhausts the ranked list. This dynamic approach adapts the search processing to the specific query, avoiding unnecessary comparisons and reducing overall system complexity while maintaining comprehensive product information.
3Loss of information
If images showing product applications and additional objects are included, then contextual information is enriched, but search relevance decreases
Solution Approach 1:
The patent applies local quality by assigning different ranks to different images based on their specific characteristics and relevance. First-rank images are identified as having high product representation quality, while second-rank images (such as application context images) are assigned lower ranks. This differentiated quality assignment ensures that contextual information is preserved in the database but does not interfere with precise product matching during search.
Data Source
Figure 1
Figure 2
Figure 3a~3f
AI summary
The present invention relates to a method of generating a ranked database, and a method of visually searching for a product in a ranked database. After determining a reference image being representative of a respective product linked to a respective product data set, for each image of the product data set a measure of similarity of the image with respect to the respective reference image is determined, and an image rank based on the determined measure of similarity is assigned to each image. After saving the plurality of product data sets comprising the image ranks to a database, a visual search for a product in this ranked database can be performed using a search image, the visual search providing search results corresponding to an image correlated to a product linked to a product data set in the database having a similarity to the search image. A threshold operation on the search results is performed, maintaining only those search results with the image rank of the corresponding image being larger than a predefined threshold.