An Analysis of Shape Based Image Retrieval Using Variants of Zernike Moments as Features

Abstract

Zernike Moments (ZMs) have been most widely used in extracting the region based shape features of an image such as Logo, Trademark, Clip-art etc. Though ZMs are assumed to be the best descriptors, still retrieval of accurate images is an important research area. In most of the researches, emphasis is given only on magnitude of ZMs and phase component is ignored. Complex Zernike Moments (CZMs) take both Phase and Magnitude component for feature extraction. Global features are extracted using ZMs or CZMs. In order to capture local details of an image, variety of options are available i.e. Wavelets, Mean and Standard deviation of Centroid distance, Curvature on edged images etc. The retrieval results of global features are compared with results obtained using combined local and global features. Experiments are performed on MPEG-7 CE-2 shape database. Recall and Precision, Bulls Eye Performance (BEP) are chosen as measures for retrieval performance. It is observed that a combination of local and global features i.e. Wavelets and Complex Zernike Moments (WCZMs) perform better in terms of retrieval accuracy without any overhead of increased feature space. Based on our experiments, WCZMs are recommended for SBIR system to achieve better retrieval accuracy. In addition to being rotationally invariant, WCZMs are found to be robust towards blur and noise. Other local features like Curvature and Centroid distance when used with ZMs prove better in terms of retrieval accuracy but WCZMs outperform this descriptor also.