• Khan, Faheem
• Ahmed, Qasim Zeeshan
• Alattal, Hessah

Abstract

Image denoising is a traditional task in image processing field and lot of research has been done on this issue. The need to improve denoising performance is a continuous challenge. In this paper, a review of the key ideas related with image denoising is presented and how this issue can be addressed using artificial neural networks as a standard nonparametric statistical tool for pattern recognition, clustering and discriminant analysis. The limitations of traditional fully connected multilayer perceptions on image processing are discussed and it is shown how we can deal with these limitations. This leads to the analysis of the currently used approach in this field known as convolutional neural networks and related Matlab toolboxes on image Processing and deep neural networks. These toolboxes use an available pre-trained denoising convolutional neural network (DnCNN). This existing framework is tested under real conditions and the outputs confirm two of the major claims behind the Matlab DnCNN: the blind denoising capabilities and low time used in the denoising task. Additionally, it was seen that the issue for low noise levels, with signal-to-noise Ratio (SNR) up to 6, the DnCNN will add more error than the noise to be removed. The last leads to suggestion that the use of DnCNN for low noise levels is worth further investigation.

Topics

• impedance spectroscopy
• clustering