Coronavirus (COVID19) is a highly contagious virus which had already killed thousands of people and infected millions more throughout the world. One of the primary challenges that medical practitioners encounter in the realm of healthcare is correctly diagnosing patients conditions and infections. So far, the gold standard screening method RT-PCR test which has been designed to detect covid-19 which only has a positive rate ranging between 30 precent and 60 percent. As a result, a system that can accurately identify images and diagnose or anticipate diseases is needed. As a result, we set out to swiftly create a compact CNN architecture capable of recognizing COVID-19-infected individuals. Different CNN architectures are suggested in this paper to extract information from X-rays which further classified into Covid-19, pneumonia, or healthy. Here, we have used two datasets from publically available repositories that are Kaggle and Mendeley [1] [2]. To see how the size of datasets affects CNN performance, we train the suggested CNNs with both the original and enhanced datasets where datasets are splitted into ratios of 80:20 and 70:30 and the comparison is shown. Also suggested CNN model is compared with the five state-of-Art pre-Trained models (VGG-16, ResNet50, InceptionV3, EfficientNetB2, DenseNet121) with the same datasets and splitting ratios. we have also used Some visualization methods through which we can get an exact idea of how CNN functions and the explanation behind the network's decisions. This study suggests a model for classifying COVID-19 patients but makes no claims about medical diagnostic accuracy. © 2022 IEEE.