Reducing Impulse Noise in Audio Signals Using Cascaded Vector Median Filtering

The interest of this paper is the reduction of random valued impulse noise in audio signals using a cascaded approach of vector median filtering. Reducing random valued impulse noise is a challenging problem because the noise corrupts the a random number of audio samples by any random number in the permissible range of signal amplitude. The vector median filter and its variants have been used effectively to reduce noise in audio and image signals. The key idea of this approach is to sequentially test and correct an audio sample using a measure of its similarity to its neighbors, the set of which is called a window. A recent method [1], which we term as “adaptive windowing median filter” proposed to use windows of varying sizes in a cascaded framework, i.e., the output of one stage is the input to the next, by using a detection and median filtering scheme. Although the filter has reported satisfactory results in reducing impulse noise, it requires intelligent choice of thresholds which is not possible always. This paper builds on the said filter and develops a cascaded vector median filter with the cascade now corresponding to the deviation of the sample to the mean of the sample-sample distance of the received signal. This idea improves on the aforementioned method as it does not require any forehand decisions and improves the filtering accuracy by virtue of effectively identifying samples that are dissimilar.