Inspecting the Efficiency of Deep Neural Networks for Classification of Images with Noise

Abstract

Noise is the undesirable data in digital images which can produce undesirable impacts like obscured objects. It becomes necessary to apply filters for denoising the images. There are many deep learning techniques for image classification. The study in this paper aims to challenge the performance of deep learning in the presence of noise in the images. The experiments are carried out on thepopular dataset- Caltech 101. The images are corrupted using Gaussian noise. Denoising is performed using combination of Median and Gaussian filters. Feature extraction methods considered are edgedetection, wavelet transformation, their concatenation, LBP and Gabor filter. Classifiers such as kNearest Neighbours, Support Vector Machine, Neural Network and Deep neural network with autoencoders are used. The experimental outcomes demonstrate that though noise can upset the execution of image mining task in terms of accuracy,deep neural networks with autoencoders has reported higher accuracy of the classification of images with noise as well as without noise.