A comparative study of dimensionality reduction methods for large-scale image retrieval

"Curse of Dimensionality" is one of the important problems that Content-Based Image Retrieval (CBIR) confronts. Dimensionality reduction is an effective method to overcome it. In this paper, six commonly-used dimensionality reduction methods are compared and analyzed to examine their respective performance in image retrieval. The six methods include Principal Component Analysis (PCA), Fisher Linear Discriminant Analysis (FLDA), Local Fisher Discriminant Analysis (LFDA), Isometric Mapping (ISOMAP), Locally Linear Embedding (LLE), and Locality Preserving Projections (LPP). For comparison, Scale Invariant Feature Transform (SIFT) and color histogram in Hue, Saturation, Value (HSV) color space are firstly extracted as image features, meanwhile SIFT feature extraction procedure is optimized to reduce the number of SIFT features. Then, PCA, FLDA, LFDA, ISOMAP, LLE, and LPP are respectively applied to reduce the dimensions of feature vectors, which can be used to generate vocabulary trees. Finally, we can process large-scale image retrieval based on the inverted index built by vocabulary trees. In the experiments, the performance of various dimensionality reduction methods are analyzed comprehensively by comparing the retrieval performance, advantages and disadvantages, computational complexity and time-consuming of image retrieval. Through a series of experiments, we can conclude that dimensionality reduction method of LLE and LPP can effectively reduce computational complexity of image retrieval, while maintaining high retrieval performance. (C) 2014 Elsevier B.V. All rights reserved.