Impact of Pixel Scaling on Classification Accuracy of Dermatological Skin Diseases Detection

Abstract

Images are made up of many features on which the performance of the system used in processing them depends. Image pixel values are one of such important features which are often not considered. This study investigates the importance of image preprocessing using some calculated statistics on the pixels of skin images in classifying images using HAM10000 dataset. Image pixel values make a great impact on the classification performance of Convolutional Neural Network (CNN) based image classifiers. In this study, the ‘original pixel values’ of the skin images are used to train three carefully designed CNN architectures. The designed architectures are further trained with some calculated statistical values using ‘global centering’, ‘local centering’, ‘dividing pixel values by the mean’ and ‘root of the division’ techniques of data normalization. The results obtained have shown that, out of the five different forms of values used in training the architectures, the CNNs trained with the original (unscaled) image pixel values perform below those trained with calculated statistics that are computed on the image pixel values