Effect of image degradation on performance of Convolutional Neural Networks

The use of deep learning approaches in image classification and recognition tasks is growing rapidly and gaining huge importance in research due to the great enhancement they achieve. Particularly, Convolutional Neural Networks (CNN) have shown a great significance in the field of computer vision and image recognition recently. They made an enormous improvement in classification and recognition systems’ accuracy. In this work, an investigation of how image related parameters such as contrast, noise, and occlusion affect the work of CNNs is to be carried out. Also, whether all types of variations cause the same drop to performance and how they rank in that regard is considered. After the experiments were carried out, the results revealed that the extent of effect of each degradation type to be different from others. It was clear that blurring and occlusion affects accuracy more than noise when considering the root mean square error as a common objective measure of the amount of alteration that each degradation caused.