Certain investigations on removal of mixed noise in digital images using cascaded convolutional neural networks

Abstract

In real-time signal processing, noise is considered an unpredictable, random phenomenon that constructively or destructively affects the pixels in an image. There are many classical filters for eliminating Gaussian Noise (GN), Salt and Pepper noise (SPN), and Speckle Noise (SN) in digital images. However, the denoising efficiency of classical filters is unsatisfactory for filtering mixed noise models like Gaussian plus Salt and Pepper noise and Gaussian plus Salt and Pepper noise plus Random valued impulse noise due to the high degree of uncertainty in the noise distribution. Recent advancements in deep neural networks helps to address these challenges in image denoising. This thesis focuses on removing mixed noises in digital images using a cascaded Convolutional Neural Network (CNN) model taking fascinating inspirations from the Visual Geometry Group (VGG) network. newlineThe proposed denoising model has a Pseudo Convolutional Neural Network (PCNN) cascaded with a Modified Convolutional Neural Network (MCNN). The proposed denoising method has two phases. In phase 1, the mixed noise corrupted images are initially processed by the pseudo-CNN. The PCNN is a typical preprocessing filter with exceptional impulse noise suppression ability in the complex noise situations. The Pseudo CNN has adaptive kernel filters whose coefficients are initialized by the noise probability function. The PCNN has a single frontward pass to calculate the best pixel estimate for the corrupted intensity values. There is no backpropagation to update the weights of the feature detectors. Hence, this network is named as Pseudo CNN. The filter coefficients of the PCNN get iv updated as the feature detector slides over the receptive fields of the noisy input image. newlineIn phase 2, the preprocessed images are fed through a Modified Convolutional Neural Network (MCNN) to obtain final denoised image. The MCNN is customized convolutional neural network inspired from the standard VGG network. To achieve improved network performance and flexibility, residua