Publication

Abstract : Obtaining neonatal thermal image data is a pricey expedition. The situations of both patients and hospitals may not be in favor of acquiring thermal images of neonates. This will leave the dataset to be an ill-representation of reality which affects the learning capabilities of algorithms. Data augmentation is one of the widely used techniques to create data that can be used when we have an imbalanced dataset or sensitive data points. Data Augmentation can be done using image translation, histogram equalization, zooming, rotation, flip, and scale called traditional methods. The diversity gained by making use of traditional modifications are relatively small when compared to DCGAN. Even though an effort is made to increase the size of the dataset using traditional augmentation methods the resultant images are still the copy of their original. We make use of DCGAN to understand and create thermal images which resemble original images yet diverse from original neonatal thermal images.