Image Similarity Search Using Compounded Feature Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image similarity search techniques in electronic commerce lack accuracy and speed in analyzing complex data, leading to low-quality search results.
Innovation Solution
An information processing apparatus that generates a compounded feature vector by concatenating feature vectors from multiple learning models, including first and second feature predictive models, a gender predictive model, and a color predictive model, to enhance the accuracy of similar image search.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single learning model is used to generate feature vectors for image similarity search, then the device complexity is low, but the search accuracy is insufficient
Solution Approach 1:
The patent segments the feature extraction process by employing multiple specialized learning models (first feature predictive model for upper-level classification, second feature predictive model for lower-level classification, attribute predictive model, and color predictive model) instead of a single monolithic model. Each model focuses on specific aspects of image analysis, thereby improving overall search accuracy while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent merges the outputs of multiple learning models into a compounded feature vector by concatenating feature vectors from different models (upper-level classification features, lower-level classification features, attribute features, and color features). This integration combines the strengths of each specialized model to achieve superior image similarity search accuracy compared to using any single model alone.
2Measurement precision
If multiple learning models are used to generate comprehensive feature vectors, then the search accuracy is improved, but the processing time increases
Solution Approach 1:
The patent implements preliminary action by pre-training and storing multiple specialized learning models before the actual similarity search operation. The models are prepared in advance to perform specific tasks (upper-level classification, lower-level classification, attribute prediction, color prediction), enabling faster inference during the search process compared to using a single large model that would need to process all aspects sequentially.
Solution Approach 2:
By segmenting the feature extraction into parallel specialized models, the patent enables simultaneous processing of different image aspects. The concatenation of feature vectors from multiple models can be performed efficiently, and the modular structure allows for optimized computation and potential parallel processing, reducing overall processing time while maintaining comprehensive feature analysis.
3Productivity
If simple feature extraction is used, then the processing speed is fast, but the analysis capability for complex data is insufficient
Solution Approach 1:
The patent changes the parameters of feature representation by generating multi-dimensional feature vectors from multiple learning models instead of using simple single-dimensional features. The compounded feature vector combines upper-level classification features, lower-level classification features, attribute features, and color features, creating a rich parameter space that enables accurate analysis of complex image data while maintaining processing efficiency through specialized model architectures.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
An information processing apparatus (1) includes: an acquisition means (101) for acquiring an object image that contains a target object; a generation means (102; 103; 104; 105) for generating a plurality of feature vectors for the object by applying the object image to a plurality of learning models; a concatenation means (106) for concatenating and embedding the plurality of feature vectors into a common feature space to generate a compounded feature vector in the feature space; and a search means (107) for searching for a similar image that is similar to the object image, using the compounded feature vector.