Abstract:
The removal of mixed-noise is an ill-posed problem due to high level of non-linearity in the distribution of noise. Most commonly encountered mixed-noise is the combination of additive white Gaussian noise (AWGN) and impulse noise (IN) that have contrasting characteristics. A number of methods from the cascade of IN and AWGN reduction to the state-of-the-art sparse representation have been reported to reduce this common form of mixed-noise. In this the-sis, a new learning-based algorithm using the convolutional neural network (CNN) models are proposed to reduce the mixed Gaussian-impulse noise from images. The models are evaluated for both the image to image learning as well as image to residual learning techniques. The proposed CNN models adopts computationally e cient transfer learning approach to obtain an end-to-end map from noisy image to noise-free image. The model has a small structure yet it is capable of providing performance superior to that of the well established methods. Experimental results on di erent settings of mixed-noise show that the proposed CNN image to image learning based denoising method performs signi cantly better than the sparse repre-sentation and patch-based methods do both in terms of accuracy and robustness. Moreover, due to the lightweight structure, the denoising operation of the proposed CNN-based method is computationally faster than that of the previously reported methods. The proposed im-age to residual learning based densely connected denoising CNN (DCDCNN) outperforms the previous state-of-the-art CNN based denoising method. Qualitative evaluation shows that the proposed DCDCNN produces visually superior images than the traditional as well as other CNN based methods. Despite being a deeper neural network architecture, the proposed DCD-CNN can denoise in a very short time by employing GPU.
