Classification of Coronavirus Images using Shrunken Features

Saban Ozturk; Umut Ozkaya; Mucahid Barstugan

is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10.1101/2020.04.03.20048868 doi: medRxiv preprint Then, four hand-crafted features are extracted from all images. By combining these feature vectors, 78 features are obtained for each image. Then the feature vectors of 260 images consisting of 78 features are over-sampling with the SMOTE method. After this process, 495 feature vectors are created. With these feature vectors, sAE and PCA are trained, respectively. The purpose of the sAE and PCA algorithms in this study are to narrow down 78 features and obtain 20 features. Finally, SVM is trained with 495 vectors containing 20 features for classification purposes. The necessity of both image augmentation and data over-sampling arises from the depth of the unbalanced structure in the dataset. In case only image augmentation is applied, there are only two images in many classes, and almost the same images will be produced. In this case, in-class overfitting occurs. When using only synthetic data over-sampling method, synthetic performance data is obtained, and performance may remain low in real applications. For this reason, two data replication techniques are combined. author/funder, who has granted medRxiv a license to display the preprint in perpetuity.