Image Search Feature Vector Generation Using Semantic Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text-based image search methods rely on text descriptions, which provide partial or one-sided information, leading to inaccurate results due to the limitations of text descriptions, and do not effectively utilize image content for retrieval.
Innovation Solution
A method that generates feature vectors for images by extracting semantic features using a neural network model, such as ResNet-152, and compares these vectors with search terms, allowing for image retrieval based on image content without relying on text descriptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If text descriptions are used for image search, then the search process is simple and fast, but the search accuracy deteriorates due to partial or one-sided information in text descriptions
Solution Approach 1:
The patent segments the image into multiple regions and extracts features from each region separately, then combines them to form a comprehensive image representation. This segmentation approach allows the system to capture detailed local information while maintaining overall image context, thereby improving search accuracy without requiring overly complex global processing
Solution Approach 2:
The patent introduces feature vectors as an intermediary representation between images and text descriptions. Instead of directly comparing images with text, the system converts both to feature vectors in a common space, enabling more accurate similarity computation while keeping the system architecture manageable
2Measurement precision
If only text descriptions are used for image retrieval, then the system is simple to implement, but the retrieval accuracy deteriorates because text descriptions provide only partial information
Solution Approach 1:
The patent merges multiple feature extraction approaches by combining region-based features with overall image features. This merging strategy allows the system to leverage both detailed local information and global image characteristics, improving retrieval accuracy while maintaining implementation feasibility through modular feature combination
Solution Approach 2:
The patent transitions from one-dimensional text description matching to multi-dimensional feature space comparison. By representing images and queries in high-dimensional feature vectors that capture color, texture, shape, and semantic information, the system achieves superior retrieval accuracy while using established dimensionality reduction techniques to keep implementation practical
3Ease of operation
If text descriptions are extracted from image context, then the search system is easy to operate, but the search results deteriorate due to imprecision and incompleteness of extracted text
Solution Approach 1:
The patent enables the image search system to self-evaluate and self-improve by using extracted text descriptions as one of multiple features rather than the sole basis for search. The system automatically combines text-based features with visual features, allowing it to compensate for imperfections in text extraction without requiring manual intervention, thus maintaining ease of operation while improving precision
Data Source
AI summary
Methods, apparatuses and electronic devices for generating a feature vector, as well as searching methods, apparatuses and electronic devices are disclosed. The method for generating a feature vector includes: acquiring data information; extracting a semantic feature from the data information, to acquire semantic feature information; and acquiring a feature vector of the data information by using a preset function, with the semantic feature information as a parameter. The technical solution identifies picture information by recognizing semantics of image information and matching the semantics of the image information with natural language descriptions. Different from conventional image search schemes of existing search engines, this technical solution does not need to retrieve a text description of image information, but retrieves and identifies images based on the content of the image information. Therefore, results with higher accuracy may be returned compared with the existing text-based image search.