Edge Detection Using Vector Quantization And Local Entropy Measures Applied To Spectrogram Component Extraction

In this paper, an approach to detect edges of signal components in multicomponent, stationary and nonstationary signals, independent of their time or frequency support, by application of K-means algorithm and local entropy measures to the time–frequency (TF) plane is proposed. The approach is based on the use of a K-means algorithm as an edge detector rather than object detector, that can maintain continuity of the edge on stationary and nonstationary signal components. Local entropy measures are applied to select appropriate K-means class to represent the edge as well as to estimate the continuity of the edge. When common edge detectors are applied, the continuity of the edge can be compromised by different occurrences such as the lack of contrast in an image or the detector working in one direction, so different methods like smoothing filters or histogram equalisation are often used as a corrective measure. The proposed K-means based method does not require additional smoothing or equalizing methods and no blurring or loss of information is present, which results in a robust method of detecting continuous edges of the useful information content in the time-frequency plane for both stationary and non-stationary signals. Since the K-means based algorithm does not depend on the contrast present in the image, the edges can se preserved in greater percent than commonly used edge detectors. Simulated signals are corrupted by different levels of Additive White Gaussian noise (AWGN) of 0dB, 3dB and 5dB. The results show the proposed method correctly detects components of an image with a continuous edge. The level of continuity of the proposed method is compared to the level of continuity of commonly used edge detectors, where results show the proposed method outperforms common edge detectors in case of non-stationary signals.