Data Dependent Denoising Techniques to Restore Noisy Images

Abstract

Tremendous advancement in the communication technology and personal assistant devices has resulted in production of large sophisticated mobile phone devices. Many mobile devices have in built cameras that require mechanical shutters to avoid salt and pepper noise. Due to absence of these mechanical shutters most of the images acquired through mobile phone cameras are affected by salt and pepper noise. Therefore it is required to have a simple, less computationally complex and effective salt and pepper noise removal technique. Standard median filter are mostly applied for the removal of salt and pepper noise. These median filters are implemented using techniques such as standard median filter, switching median filters, adaptive median filter, and adaptive centre weighted median filter. These filters have demonstrated effective performance in removing salt and pepper noise at lower densities. However it is desired to remove salt and pepper noise at higher densities due to absence of mechanical shutters in cameras. Adaptive filter that increases the filter size during the processes of filtering of noisy images have been proposed for removal of high density salt and pepper noise. Mostly this adaptive filter starts filtering with a minimum filter size of . It periodically increases filter size by two if the condition specific to the image quality (noise removal) are not satisfied. The proposed algorithm consists of impulse noise detection and noise removal modules. An automatic impulse noise detection module is based on mean and variance technique that selects the noisy pixels among the entire image. The noise removal module is based on replacement of noisy pixel through mean and edge direction using Gabor filter. Several experiments are performed to establish the effectiveness of the proposed technique using PSNR, MBE, SSIM and IEF parameters. The image does not contain any visual defects for the noise density 0.1 to 0.5 hereafter the noise pixels for noise density greater than 0.5 are clearly visible. Also for the n