Balanced image clustering method based on hierarchical clustering

Abstract

The invention discloses a balanced image clustering method based on hierarchical clustering. According to high dimensional feature data of apparel goods images, clusters with balanced sizes are obtained by a method based on hierarchical clustering, and the quantity of data in a single cluster does not exceed a limited threshold value. When in retrieval, after distances among retrieved data and all clustering centers are computed, a plurality of the nearest clusters are selected, and the data are accessed inside the clusters to obtain final research results. Compared with a universal indexing method based on clustering, the method avoids the problem of overlarge accessed data volume when the retrieved data are in large clusters, and ensures research performances. Besides, by accessing a plurality of clusters, the research results and SSA (serial storage architecture) research results have a higher overlap ratio, and research effects are improved.

Description

technical field [0001] The invention relates to the technical field of image search, in particular to a fast approximate k-nearest neighbor retrieval method for high-dimensional image vectors based on hierarchical clustering. Background technique [0002] In content-based image retrieval technology (Content-Based Image Retrieval, CBIR), when a user uploads a product image and expects to search for products that are the same as or similar to the image, the search engine performs feature extraction on the product image uploaded by the user. And select the k images closest to it in the high-dimensional space from the indexed image feature vector database as the result and return. To query the nearest k image features in a large database of indexed features, the most basic method is the SSA method. The SSA method obtains the nearest k images by calculating the distance between the retrieved image and each stored image, and then sorting these distances. This is an exact k-Neare...

AI Technical Summary

Problems solved by technology

Therefore, if a cluster contains a much higher than average amount of data, it will seriously affect the average response time of the product image search engine

[0005] 2. Data traversal is limited to clusters. If there is k-nearest neighbor data in other clusters, it will be lost in the retrieval results, resulting in reduced query effect

Images