Co-occurrence of maximal Haar-like wavelet filters for CBIR

Abstract

This paper introduces a novel technique, co-occurrence of Haar-like wavelet filters (CHLWF), for content-based image retrieval (CBIR). CHLWF obtains correlation among dominant Haar-like wavelet filters to capture structural similarity of images. Selection of dominant Haar-like wavelet filter avoids less prominent directions of intensity variation and reduces complexities of feature computation. Most of the correlogram-based techniques need to decide quantisation thresholds beforehand but CHLWF excludes this as the quantisation is performed automatically based on directions of maximum intensity variations. Performance of proposed feature is justified through three benchmark image databases viz. Corel 1000 (DB1), Brodatz texture (DB2) and MIT VisTex (DB3). Experimental results on DB1 have demonstrated that average precision and average recall are significantly improved by CHLWF in comparison to optimal quantised wavelet correlogram, Gabor wavelet correlogram and cross correlogram (CC). Likewise, on DB2 and DB3 databases, average retrieval rate of CHLWF is better than CC, dual tree complex wavelet transform (DT-CWT), dual tree rotated complex wavelet transform (DT-RCWT) and DT-CWT + DT-RCWT etc. respectively.