Image Search Indexing via Scalar Similarity and Difference Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In image search systems, when the directions of difference vectors are opposite, the sum vector's element values become smaller, leading to inadequate reflection of image features, resulting in inaccurate searches.
Innovation Solution
An image search system that calculates scalar values indicating similarity between feature vectors and representative vectors, generates feature values based on these scalar values, and creates a search index for accurate image retrieval, using techniques like clustering and autoencoders to compress data while maintaining feature accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If difference vectors are summed to compress feature data, then data amount is reduced, but search accuracy deteriorates when difference vectors have opposite directions
Solution Approach 1:
The patent segments the feature representation into two independent components: visual words (from clustering) and difference vectors (from pairwise comparisons). This segmentation allows each component to serve its specific purpose - visual words for category-level representation and difference vectors for fine-grained discrimination - thereby resolving the contradiction between data compression and search accuracy.
Solution Approach 2:
The patent transitions from a single-dimensional feature representation to a multi-dimensional space by combining visual words with difference vectors. The difference vectors are normalized and combined with visual word embeddings in a higher-dimensional feature space, enabling the system to capture both coarse and fine features simultaneously without the cancellation problem of simple vector summation.
2Quantity of substance
If sum vectors are used to represent images, then data compression is achieved, but feature representation becomes inadequate when difference vectors oppose each other
Solution Approach 1:
The patent extracts the problematic cancellation effect by separating difference vectors from the summation process. Instead of summing difference vectors directly (which causes information loss when they oppose), the system extracts difference vector directions, normalizes them, and combines them with visual word representations. This extraction preserves directional information while avoiding the harmful summation cancellation.
Solution Approach 2:
The patent changes the parameter representation from raw difference vector sums to normalized difference vector directions combined with visual word embeddings. By normalizing difference vectors to unit length and combining them multiplicatively rather than additively, the system preserves the magnitude and direction information without the cancellation problem, thereby reducing information loss.
3Quantity of substance
If feature vectors are compressed into visual words, then data amount is reduced, but similarity measurement accuracy decreases
Solution Approach 1:
The patent merges two complementary representation mechanisms: visual words for coarse-grained similarity and difference vectors for fine-grained discrimination. The combined representation formula integrates both components, allowing the system to achieve both data compression through visual words and accurate similarity measurement through the differentiated difference vector information.
Data Source
AI summary
To search for a similar image more accurately, an image search system acquires a plurality of representative vectors, which are generated based on a plurality of feature vectors each indicating a feature of an image and included in a feature vector space, calculates a scalar value indicating a degree of similarity between each of the plurality of feature vectors and one of the representative vectors corresponding to the feature vector, calculates, for each of images, a feature value indicating a feature that depends on the representative vector based on the scalar value, and creates a search index associated with the calculated feature value.