A new HSI based filtering technique for impulse noise removal in images

Abstract

We propose a new technique for impulse noise filtering that can remove the impulse noises from color as well as gray scale images. We operate on the HSI (Hue-Saturation-Intensity) color model. Our algorithm has three Phases. In first Phase, we take a window W of size N×N (say, 3×3) and form two groups: group of color and group of colorless pixels. We select the group that has the higher count of pixels in W. This allows us to remove the noise due to the colorless pixels from the color pixels and vice-versa. In the second Phase, if the selected group is a collection of colorless pixels then we find the median pixel based on increasing order of Intensity values and we call this as a candidate pixel. If the selected group is a set of color pixels then we will convert the 3-dimensional pixels into 1-dimension by first sub-grouping them based on their Hue values and selecting the sub-group that has maximum count. The pixels in this selected sub-group are homogeneous in nature with respect to the color and they vary by their Intensity and Saturation values. We find the average of the Intensity values and select the candidate pixel which is near this average. The Saturation, optionally, can be used to break the tie between the pixels that claim, at the same time, to be nearest the average. In third Phase, we decide whether the center pixel, in the current window W, is a noisy or noiseless based on an adaptive threshold which depends on the absolute difference of the Intensity, and/or Hue of the center pixel as that of the candidate pixel. If we find the center as noisy then the candidate will replace it or else we leave this center as intact. We have compared our method with the standard vector median filter (VMF), used for removing impulse noise from color images. The experiments tell us that our method gives better result with respect to the quality of the image (visual appearance), time for computation, and removal of noise. We present the results on a few color and gray scale images that are corrupted by salt and pepper noise to demonstrate the effectiveness of our approach.