Multiresolution Analysis using Complex Wavelet and Curvelet Features for Content based Image Retrieval

Abstract

In a typical content-based image retrieval (CBIR) system, retrieval results are a set of images sorted by feature similarities with respect to the query image. This paper demonstrates the comparative study of retrieval performance of CBIR system using real dual-tree DWT (R-DT-DWT), complex dual-tree DWT (C-DT-DWT) and Curvelet Transform. The experiments are carried out on Corel database of 1000 images database of 10 different classes with various similarity measures. The overall performance for Canberra distance was found to be better as compared to Minkowski and Manhattan distances. Experimental results indicate that the proposed method gives excellent average precision of 100% for Dinosaur class and 95% for roses class of images. Comparing the results and taking feature vector size into consideration, it may be better to opt for R-DT-DWT rather than C-DT-DWT or Curvelet features for feature extraction. But curvelet features contains more directional information at high frequencies and high frequency components provides better discrimination between images.