Feature extraction for content-based image retrieval

For content-based retrieval, three image features, namely, Zemike moment invariants, Fourier descriptors, and wavelet descriptors, are studied. They are extracted from shapes of constituent semantic entities or objects and are then compared with image queries posed in the form of sketches. To measure similarity between an image candidate and a query, the Euclidean distance with corresponding weights is employed. For the wavelet descriptors, a new method for finding a starting point is proposed. On the basis of the results obtained, and for the images tested, this study has shown that the wavelet descriptors provide the most effective features. The performance of the Zemike moments is at the end. The recognition rate of the descriptors is 100 % for all queries derived via affine transformation from the images in the database. For deformed queries posed as free-hand sketches, the wavelet descriptor correctly recognises four cases out of five.