A COMPREHENSIVE EXPERIMENTAL STUDY FOR EVALUATING THE PERFORMANCE OF WELL-KNOWN CNN PRE-TRAINED MODELS IN NOISY ENVIRONMENTS

Abstract

This work aims at testing the efficiency of the pre-trained models in term of classifying images in noisy environments. To this end, three different well-known pre-trained models i.e. MobileNet, ResNet, and GoogleNet have been tested using a high-quality image dataset, then the model has been saved to be tested after injecting some noise to the images. Particularly, we proposed injecting Gaussian noise into the images in the original dataset gradually to see how the performance of that models can affect with respect to the proportion of the noise in the image. We noted that the performance of the model dropped down dramatically in accordance with the noise proportions, where the classification accuracy of the tested models dropped down from 98.66, 99.00, 98.00% for MobileNet, ResNet, and GoogleNet respectively when the original dataset has been used to 35.33, 73.33, 68.99% when 75% noised dataset has been used for testing the models. To the best of our knowledge this is the first time that the effects of noise on deep learning models have been experimentally studied. The obtained results reflect the need for some auxiliary models that should be used as a pre-processing phase to improve the performance of these models in order to be applied in real-life applications.