Some Novel Approaches Towards Improving Noise Reduction And Effective Classification Of Satellite Images

Abstract

The primary objective of this thesis is to develop methods for reducing random valued impulse noise in satellite images. The conventional filtering methods function by applying a noise reduction scheme, typically the vector median filtering approach and its variants, to the central pixel of a suitably selected window that slides iteratively across the entire image. This thesis proposes two novel approaches to reduce impulse noise in satellite images. The first approach is Modified adaptive switching trimmed vector median filter with Non Local Means Integration. The proposed algorithm initiates denoising with Non-Local Means, utilizing patch-based similarity measurements within a designated search window. This non-local strategy guides the denoising process based on patch correlations. Subsequently, the approach transitions to Adaptive Switching Trimmed Vector Median Filter, utilizing local noise estimation factors to intelligently select between vector median and trimmed median filtering. This adaptability effectively addresses varying noise characteristics, ensuring noise reduction without compromising essential image structures. The seamless integration of Non-Local Means and Adaptive Switching Trimmed Vector Median Filter leads to denoised images with reduced artifacts, preserved details, and improved perceptual quality. The Second approach proposes a novel method for denoising satellite images using Isolated Vector Median filtering with k-means Clustering (IMF-KM). The key concepts of the proposed IMF-KM method are (a) classification of pixels in the sliding window into those contributing to the signal and those contributing to the noise, and (b) to perform median filtering, in isolation of the color components, over the signal-contributing pixel intensities. Using the k-means clustering algorithm, step (a) categorizes the pixels based on the spatial placement of their intensities. The median filter is then used to process the cluster of intensities that contribute to the signal. This process is done on each