LOCAL AND GLOBAL DESCRIPTORS USING WAVELET BASED FEATURES FOR IMAGE RETRIEVAL

Abstract

This research work develops a content based image retrieval system using local as well as global color and texture features. In content based image retrieval, most important task is to collect features and once we have these features next task is to combine these features for achieving good retrieval efficiency. To address this problem, six new techniques are presented. Discrete wavelet transform (DWT) has suffers with two main disadvantages i.e. lack of shift invariance and poor directional selectivity. First problem of standard wavelet transform is overcome using two-dimensional rotated wavelet filters (RWF) which captures diagonal information more accurately when compared to discrete wavelet transforms (DWT). In most of the CBIR systems, mean, standard deviation and histogram of the transform coefficients are used as the global feature vectors. But in proposed method, combinations of spatial orientation tree (SOT) and vocabulary tree (VT) are used for feature collection which is useful to collect the local feature of image in transform domain. Second problem of DWT is conquering using dual tree complex wavelet transform (DT-CWT) and dual tree rotated complex wavelet transform (DT-RCWT) in combination with spatial orientation tree (SOT) and vocabulary tree (VT) with additional directional information. In CBIR, speed is also play very important role in addition to retrieval efficiency and these goals are achieved using a trous wavelet transform (AWT) with texton concept. AWT gives shift invariant wavelet features and texton elements give spatial relation in horizontal, vertical, diagonal and minor diagonal directions. Final feature vector is generated by performing co occurrence operations on texton image in horizontal and zigzag direction. For integration of two features, first attempt is made using the combination of global and local features. Global feature is collected from color histogram (CH) and block bit-plane (BPP) is used for local feature collection. From the experimental analysis it is clear that the individual performance of CH and BPP is not up to the mark but it shots up once used in combination. Similarly, the problem of assigning the weights to individual feature is solved by particle swarm optimization technique. Color histogram is well known method in image retrieval application but it has two main limitations i.e. it gives only global information and spatial information is completely ignored. First problem is solved using the local color histogram in combination with vocabulary tree. Further, this color feature is integrated with texture feature which obtained from the combination of DWT, RWF, SOT and VT. Second problem of color histogram is attempted using interspaces correlogram scheme which collects the color features at spatial level using interspaces correlogram technique. Similarly the texture features are collected at local level using the combination of a trous wavelet transform (AWT) with local binary pattern (LBP). Further, the combination of these color and texture features are utilized for achieving better performance in image retrieval application.