Compact feature learning for multimedia retrieval

Abstract

Efficient multimedia search has attracted much attention in recent years due to the exponential growth of data available. Extracting representative compact features for nearest neighbor (NN) search and approximate nearest neighbor (ANN) search has played an important but challenging role in multimedia retrieval systems. This is because efficient compact features require two essential properties: (1) It should be able to capture high quality information from raw data; (2) It needs to be of small dimensionality to support fast search with low memory costs. While several compact feature learning algorithms have been proposed in the literature and some of them have achieved reasonably good performance in retrieval benchmark datasets, there is still some room for further improvement. Hence, this thesis is dedicated to developing several compact feature learning algorithms for different multimedia search systems using machine learning concepts. Particularly, we present four compact feature learning methods. Experimental results in benchmark retrieval datasets and comparisons with popular feature learning and hashing methods have demonstrated the effectiveness of our proposed methods.