Osteoarthritis (OA) is a major cause for mobility impairment, specifically among women. Due to lack of medical facilities and expertise in the remote areas, OA detection occurs at quite severe stages when already it has started affecting the mobility and the recovery is difficult. OA severity is usu-ally measured through Kelligen-Lawrence (KL) grades. Simple radiography (X-ray imaging), being non-invasive, cost-effective and easily available, is considered an important tool for early detection and mass scanning. But, it is less accurate. The state-of-art literature shows that the accuracy obtained on OA severity detection using radiograph has a better scope of improvement. It is expected that deep learning will provide a better accuracy given good amount of training data. Also, most of the work done till now is on the datasets from the western subjects. Due to the difference in knee structure, it is expected that the systems developed for OA severity stage detection on western subjects will not be equivalently accurate for Indian subjects. Accordingly, the current article targets deep learning based OA severity detection with better accuracy for Indian subjects. It successfully achieves it by employing transfer learning through EfficientNetB1 with the test accuracy of almost 89% on the database from Indian subjects.