Graph Structural Feature Image Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the number of images stored in databases increases, accurately searching for desired images of a specific category becomes increasingly difficult due to the vast quantity of data, leading to inefficiencies in image retrieval processes.
Innovation Solution
The method involves generating graph structural features from both input and target images by encoding text content and relative position relationships, using convolutional neural networks to determine similarity, and outputting matching images based on a threshold comparison of these features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the quantity of images stored in the database is increased, then the coverage and comprehensiveness of the image database is improved, but the difficulty and time required to search for desired images increases
Solution Approach 1:
The patent applies preliminary action by pre-processing images during the database population phase. Graph structural features are extracted and encoded into feature vectors in advance, and an inverted index is pre-built mapping feature vectors to image identifiers. When a search query arrives, the system encodes the query image's graph structural features and directly retrieves candidate images using the pre-built inverted index, avoiding the need to compare the query against all images in the database, thus significantly reducing search time despite the large database size.
2Quantity of substance
If the quantity of images stored in the database is increased, then the coverage and comprehensiveness of the image database is improved, but the accuracy of image search results deteriorates
Solution Approach 1:
The patent applies parameter changes by transforming the image representation from raw pixel data to graph structural features. Images are converted to directed graphs where nodes represent text blocks and edges represent spatial relationships, then encoded into feature vectors using graph neural networks. This parameter transformation enables the system to capture semantic and structural information that is invariant to image size and layout variations, maintaining high search accuracy even when searching through large quantities of diverse images.
3Device complexity
If traditional image search methods are used on large databases, then the simplicity of the search system is maintained, but the effectiveness and precision of image retrieval deteriorates
Solution Approach 1:
The patent introduces an intermediary layer between the query image and the image database. Graph structural features serve as the intermediary representation that bridges the query and stored images. The system extracts graph structural features from both the query image and database images, encodes them into feature vectors, and uses these vectors for comparison and retrieval. This intermediary representation enables effective image retrieval by capturing essential semantic and structural information while filtering out irrelevant variations.
Data Source
AI summary
Embodiments of the specification provide an image search method, an apparatus, and a device. The method includes: obtaining an input image associated with an image search, wherein the input image includes a plurality of first text blocks; selecting a to-be-processed image from a target database, wherein the to-be-processed image includes a plurality of second text blocks; and generating a first graph structural feature based on the plurality of first text blocks; generating a second graph structural feature based on the plurality of second text blocks; determining that the first graph structural feature and the second graph structural feature satisfy a condition; and in response to determining that the first graph structural feature and the second graph structural feature satisfy the condition, outputting the to-be-processed image as a search result.