A non-probabilistic recognizer of stochastic signals based on KLT

This paper presents an efficient algorithm which is able to accurately recognize non-deterministic signals generated by synthetic non-chaotic and chaotic stochastic processes (SPs), as well as by natural phenomena (that are inherently stochastic) such as speech, image, and electroencephalographic signals. This recognition algorithm exploits a Karhunen–Loève transform (KLT)-based model able to characterize signals in terms of non-deterministic trajectories and consists of the concatenation of (i) a training stage, which iteratively extracts suitable parameter collections by means of the KLT and (ii) a recognition procedure based on ad hoc metric that measures the trajectory-proximities, in order to associate the unknown signal to the SP which this signal can be considered a realization of. The proposed methodology is able to recognize SPs without estimating their probability density function (pdf), thus requiring a low computational complexity to be implemented. Exhaustive experimentation on specific case-studies was performed and some experimental results were compared to other existing techniques such as hidden Markov model (HMM), vector quantization (VQ), and dynamic time warping (DTW). Recognition performance is similar to current best practices for non-chaotic signals and higher for chaotic ones. A better noise rejection was also achieved, and a reduction of two orders of magnitude in training-times compared with HMM was obtained, thus making the proposed methodology one of the current best practices in this field. Finally, the experimental results obtained by three different applications of the recognizer (an automatic speech recognition system, an automatic facial recognition system, and an automatic diagnosis system of the ictal and interictal epilepsy) clearly show excellent classification performance, and it is worth noting as complex filters are not needed unlike other current best practices.