Design, implementation and performance evaluation of similar image retrieval system based on self-organizing feature map

摘要

The paper describes a new method to extract and cluster image features for effective still image database. The features concerning color and texture are extracted using the multiresolution analysis. Contrast to traditional image databases where feature vectors extracted from stored images are stored and are used to match the feature vector of the input image1 we use the Self-Organizing Maps neural network for clustering stored images and generate topological feature maps with codebook vectors represented similarity between feature vectors. No feature vectors is stored in the databases, since similar retrieval is performed between codebook vectors and feature vectors. A prototype image database is developed and we experiments on the method of retrieval by example and subspace for image data. The paper reports on the architecture and experimental results.