An enhanced normalisation technique for wavelet shape descriptors

Wavelet shape descriptors are widely used in shape recognition and retrieval due to their multi-resolution nature and ability to maintain local shape features. Unfortunately, this extra information creates significant problems in the generation of a suitable normalized descriptor for shape matching. Several techniques have been proposed to alleviate this problem, but all have weaknesses which decrease the discriminating capacity. In this paper, a new combined strategy is proposed, which uses shape normalization prior to wavelet processing for translation and scaling, and descriptor normalization for starting point and rotation. Shape reconstruction ability and retrieval capabilities are assessed experimentally and compared to existing approaches using a small database of shapes derived from trademark retrieval research. The combined normalization technique is shown to produce a descriptor that is more perceptually aligned, and hence more accurate for retrieval tasks. In addition this process is lossless; hence the original shape can be perfectly reconstructed.