Concise Comparison of CNN Models On a Specified Dataset

Abstract

Recently, interest in Deep Learning (DL), which is a subset of Machine Learning (ML), has emerged. The most famous and used from the DL is the Convolutional Neural Network (CNN). CNN is particularly effective in image processing. There are many duties in image processing that CNN can do, i.e., segmentation, classification, object detection, facial recognition, etc. Image classification is one of the most important applications due to its relevance to various fields, including the healthcare industry and others. One of the challenges researchers face is selecting the appropriate algorithm for the classification task, particularly when dealing with binary or multi-class classification. This paper attempts to compare these algorithms depending on a specific dataset which have four classes, each having a balanced number of medical images. The main thing that the paper focuses on is the power of these algorithms in image classification when placed in the same conditions. This paper also makes a comparison inside the model itself by using three scenarios. The first one involves binary classification, the second uses three classes from the dataset, while the third scenario uses the entire number of classes. The best result among the models is going to AlexNet with an accuracy of 91.92%, and the DenseNet169 with an accuracy of 91.48%. Finally, this paper highlights the differences among state-of-the-art algorithms, particularly in their application to binary and multi-classification tasks.