The Study of Convid-19 Recognition and Performance Optimization Based on AlexNet

Abstract

Corona Virus Disease 19 (COVID-19), a severe disease that killed millions of human’s lives. It has been found a quick and highly efficient way to classify whether a people get COVID-19 or not. Computer Vision (CV) is used to scan the chest X-ray image and help identify patient’s condition. In this paper, one of deep learning algorithm - AlexNet provides several hyper-parameters to increase the system. Four main hyper-parameters are modified to increase the accuracy: size of input data, epochs, learning rate and Batch Normalization (BN) layer. In the first experiment, compare the accuracy of two input data size. Resizing is a crucial process in this test. Since the default image size of AlexNet is 256, it is required to change the size in both training and test parts. Then, three learning rate are tested to find which one is the steadiest one. ReduceLROnPlateau (RLROP) is one of the best learning rate methods to AlexNet. It can modify the number every time if the accuracy doesn’t increase. BN layer gives a surprised result to the classification report: it has a negative effect to the system and has been denied quickly. During the two experiments, the value of epochs is constant. So, it is easy to find the best epoch number. The consequence of classification report shows that a smaller image size without BN layers gives a highest accuracy and has a highly efficient and steady system.