The knowledge of artificial intelligence, especially, machine learning has been widely used to handle various real-world applications and related societal challenges. One of the crucial issues faced by visually impaired people is to correctly recognize paper currency and its denomination easily. The paper proposes a convolutional neural network-based currency recognition system applied to the Indian Currency. The basic features include data pre-processing to filter out the old currency notes and incorrect labels, followed by a pre-trained model selection. The proposed model is trained using transfer learning, and afterward, a custom layer is added to recognize currency notes. Further, a fully connected customized neural network architecture has been designed. The model undergoes fine tuning of essential hyperparameters. The proposed model classifies Indian banknotes with more reliable accuracy (99.5%), which is even flexible to different image orientations (flipped, folded, out of frame, etc.) as well as empty backgrounds for a real-time application.