Augment image data using noise

Abstract

Convolutional Neural Network (CNN) models for image classification have made strides in various fields such as identifying diseases in the medical industry. However, their performance is greatly affected by the data used to train them. Factors such as data quality, generalized data, and data quantity contribute to the performance of CNN models. Noisy, esoteric, and too little data will decrease the classification accuracy of these models. This paper studies the effects of noise augmentation, a method to combat the negative factors above, by injecting various noise types into an X-ray image dataset. The noise types used are Rayleigh, Uniform, Laplace, and Negative Exponential distributed noise. A pre-trained DenseNet121 model is used to conduct training and testing. The metric evaluations highlighted here are Loss, area-under-curve (AUC) score, peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM). By comparing models trained with the dataset of different noise types using Loss and AUC score, we determined the most suitable noise type here to be Rayleigh distributed noise. Further model testing is done using Rayleigh noise-augmented dataset and original-trained dataset for observation along with similarity tests of the 2 datasets. This resulted in both models performing similarly based on their average AUC scores of around 0.75, and it is reflected in the PSNR and SSIM tests as well.