An Experimental Analysis on Effect of Noisy Image Datasets on Performance of Deep Learning Models

Abstract

Deep Learning is a sub field of Machine Learning which works based on neural network structures. Deep Learning has been increasing the capabilities of Artificial Intelligent systems rapidly in many emerging areas like video analytics, data analytics and autonomous systems. The availability of the labelled data is increased due to enhancements in digital technology. The generation and labelling of training data become easier as many advanced and low-cost sensors are developed. Even though various sensors are available, Noise is the one of the major concerns of the database creation. In case of limited availability of data, Noise can be used to normalize the model to avoid overfit. But in huge databases, noise decreases the inference accuracy of the trained model. So, it is very important to understand the effect of noise on model performance. This paper presents an experimental analysis on the individual and combined effect of noise present in training database and validation database. Here, Standard CIFAR-10 database with 60000 images is used as labelled data. A comparative analysis is made to observe the behavior of trained model in terms of inference accuracy against three types of injected noises such as Gaussian Noise, Salt & Pepper Noise and Speckle Noise at different variance levels.