Classification of Bangladeshi Currency Using Convolutional Neural Network in Cross-Dataset Recognition Environment

Bangladeshi banknotes denominations have evolved in recent years and have made classification of notes with different designs and colors, a challenging task. There is a scarcity of an appropriate dataset with a significant amount of Bangladeshi banknotes and a wider spread of variations in terms of position, lighting, and background. This paper utilizes a transfer learning approach to recognize Bangladeshi banknotes with distinct hyper-parameters from four pretrained CNN models, namely VGG16, VGG19, InceptionV3, and Inception-ResNetV2. Using such large models can be very expensive in terms of computation, storage, and time, which can be handled by large-scale organizations. For smaller institutions, high-cost machines using a large model might be a budget constraint. Hence, this paper also proposes a minimalist CNN model to effectively train with a relatively large dataset that takes a short training period and lower computational cost. The pre-trained models and our proposed model were trained, validated and tested with a dataset of 20,000 RGB images. We introduced cross-dataset recognition test on all the models and compared their generalizing ability in terms of accuracy rates. For cross-dataset recognition, we have created a dataset of 1,446 RGB images considering different angles, positions, backgrounds and tints of the banknotes. We have investigated the four most commonly used Bangladeshi banknotes: 50, 100, 500, and 1000 Taka. Our proposed model outperformed the state-of-the-art pre-trained models with an accuracy of 70.47% on the cross-dataset recognition test.