Abstarct: COVID-19 is a respiratory disease that affects the lungs and can lead to respiratory problems and severe pulmonary infections. The virus, identified in late 2019, has claimed millions of lives. Despite the development of vaccines and improved treatment methods, the risk of contracting the virus remains, especially for individuals with underlying medical conditions. Accurately diagnosing the severity of COVID-19 is crucial for providing essential services to patients, improving treatment outcomes, and reducing complications and mortality associated with the virus. Proper assessment of disease severity can lead to optimal allocation of healthcare resources, selection of appropriate treatment methods, and closer monitoring of patients. However, determining the severity of COVID-19 is challenging, as distinguishing between mild cases and Normal-PCR+ is difficult. COVID-19 is classified into four severity categories: Normal-PCR+, mild, moderate, and severe. To improve the diagnosis of COVID-19 severity, this thesis introduces a novel method called DA-COVSGNet. In the proposed method, X-ray images are first preprocessed using SMOTE and CLAHE techniques, and then these images are fed into the feature extraction section of a convolutional neural network. New attention mechanisms are also employed to enhance the differentiation and classification of disease severity levels, leading to higher accuracy in identifying severity categories. These mechanisms allow the model to focus on important regions of the image and extract relevant information more accurately. Finally, the proposed method was evaluated on the COVIDGR dataset. Results show that the proposed method achieved accuracies of 96.7%, 96.2%, 98.5%, and 95.2% for the Normal-PCR+, mild, moderate, and severe categories, respectively. These results represent improvements of 36.7%, 22.5%, and 7.1% in the Normal-PCR+, mild, and moderate categories compared to the previous best method, AANet.