A Performance Evaluation of Deep Learning-Based Image Classification Models on Noisy and Augmented Datasets

Image classification on several datasets have been attempted by many researchers using one model or the other. These classification tasks using different models work effectively in many cases when the input images are clean in the sense the images are without noise, not blurred, and free from clutter. However, the performance of the classification models degrades when the input images are noisy, blurred or contain clutter. In this paper, the performance comparison of various classification models is done using clean images as well as noisy and augmented images. Different types of noise such as gaussian, Salt and prepper noise is added to images and classification is done on the original images as well as on the noisy and augmented images. The experimental results show that the ResNet50 performs better than VGG16 in terms of robustness to noise. The robustness index to evaluate the robustness of the models for noisy images is also discussed.