Automatic recognition of noisy digital images using Deep learning

Abstract

In the field of computer vision, image classification is one of the main problems that hold the lion’s share of research. Since in most real-world scenarios within image classification applications there is no control over how the qualities of the images are given. Remarkably, it is crucial to consider that these images might be damaged by noise intentionally or unintentionally. In this thesis, we try to clarify the effects of noise contained in the images in any way possible within image classification tasks by analyzing two different types of noise (S&P, Gaussian) with five different levels on three CNN models (XceptionNet, GoogleNet, ResNet) using the same parameters (Dataset, noise, and level of noise), and how denoising methods can help to alleviate this problem. We perform our experiments with the Cifar10 dataset and two different denoising methods (one for each type of noise), our results show that noise in images can hinder classification tasks and cause it a problem (make it harder to separate classes). Although images were denoised, we were unable to reach the results obtained in the noise-free scenarios.