Knee osteoarthritis (OA) is a common musculoskeletal disease. Nowadays, OA is identified by looking at the symptoms and evaluating plain radiographs. The disease cannot be predicted in the early stage through such methods. Identifying the disease by evaluating the radiographs accurately and predicting the severity level of the disease is not an easy task. In this technology era, technology in the health sector is used to treat patients effectively. Advanced prediction models such as machine learning should be used to classify the patient accurately based on the severity level of the disease. Convolutional Neural Network (CNN) predicts the knee OA severity using the Kellgren-Lawrence (KL) grading. The five different classes, standard, severe, mild, moderate, and doubtful, are classified using standard CNN models. The confusion matrix is plotted for these five classes, and its values are plotted and analyzed. The web application was designed using Flask to identify the severity level in the proposed system. Visual Geometry Group (VGG-16) is also explored in this research, including the limitation of disk space and estimated bandwidth with gradient size. The three models, i.e., standard CNN, VGG-16, and DenseNet-121 accuracy, are calculated in that DenseNet has the highest accuracy of 78%, with a testing loss of 0.85. The DenseNets need few parameters and allow feature reuse, achieving state-of-the-art performances and providing better results with this dataset. This DenseNet can overcome the limitations of VGG-16 models, ResNET, and Standard CNN models.