Publication

Abstract : Transfer Learning-based Deep Learning is the most popular technique adopted in biomedical image classification. In the first step of this approach, a pre-trained convolutional neural network (CNN) is used as a feature extractor. The second step involves fine-tuning the hyperparameters, namely learning rate, batch-size, and epoch-size. However, due to large number of input images and large image size, the CNN demands high computational cost and huge training time. This chapter on sparse image classification adds a compression layer before the input images are fed to the classifier. The compression layer sparsifies the images in the transform domain. In the cases of biomedical images, the wavelet domain provides a sparse basis for representation. Thus, images are compressed by a significant factor known as the compression ratio (CR) without significant loss of image quality determined by Percentage Root Deviation (PRD) according to the theory of approximate computing. In this paper, classification of images belonging to four different brain disease, viz., Endema, Necrosis, Enhancing Tumor, Non-Enhancing Tumor, has been performed by sparse image classification of MRI images.