Real-time image and signal recognition with the discrete rectangular wave transform

Douglas Yong-shik Kim, Purdue University

Abstract

This thesis deals with real-time image and signal recognition based on the discrete Rectangular Wave Transform (DRWT) as well as scale-invariant recognition and scale estimation based on the discrete Rectangular Mellin Transform (DRMT). The methods proposed have the potential to drastically decrease the time of computation and cost of implementation while they also usually increase the quality of recognition and scale estimation, especially in the presence of noise. Both software and hardware issues for the real-time implementation of the recognition tasks can be handled with the DRWT and the DRMT methods. The DRWT is a very simple transform, whose matrix elements consists only of combinations of $\pm 1, 0, \pm$j. It is the preprocessing part of the discrete Fourier transform (DFT) in the two-stage representation of the DFT. The DRWT is mainly utilized for the purpose of feature extraction followed by a classification algorithm. In all the experiments carried out so far in shape recognition, signal recognition and image recognition, the DRWT usually performed better than the DFT in terms of classification accuracy. Since its computation involves only elementary operations, it is much simpler, and thereby considerably faster and easier to implement in VLSI and electro-optical architectures as well as in microprocessors than DFT. These conclusions are even more true in the case of the DRMT in comparison with the discrete Mellin transform and its fast implementation since the FFT techniques are more difficult to use in this case. In this thesis, both experimental and theoretical work is discussed to explain these results.

Degree

Ph.D.

Advisors

Ersoy, Purdue University.

Subject Area

Electrical engineering

COinS