FPGA implementation of deep learning model utilizing different normalization algorithms for COVID-19 diagnosis

Abstract

Normalization is utilized to remove outliers from the dataset and address network bias. In this research, Mean-Variance-Softmax-Rescale (MVSR) and Min-Max normalizations are employed in various combinations for the diagnosis of COVID-19 using a Convolutional Neural Network (CNN)-based Deep Learning (DL) model, aimed at enhancing network accuracy. To accomplish this, the CNN model is developed within the Google Colab environment and trained using a publicly available dataset consisting of chest X-ray images related to COVID-19. The dataset is normalized using different combinations of the MVSR and Min-Max normalization algorithms to compare model accuracy. Each normalized dataset is used for model training, and subsequently, each trained model has been saved as a .h5 file and loaded into the Kria KV260 Vision AI Starter Kit FPGA for the testing phase. The most accurate results are obtained when MVSR and Min-Max normalizations are applied simultaneously. This high-performing scenario is re-evaluated with COVID-19 and normal X-ray images on FPGA configuration. Experimentally, the highest accuracy is achieved in real-time with the MVSR+Min-Max scenario, reaching 93%. The model's precision, recall, and F1-Score values are determined as 0.91, 0.96, and 0.93, respectively.