Bangladeshi Food Classification Using Transfer Learning and Ensemble Model

Automated food intake monitoring has been deemed an imperative pursuit. Despite the introduction of various datasets and deep learning techniques, there has been a growing demand for such systems, giving rise to challenges in keeping up with the multitude of food types across different countries. To tackle this issue, the authors developed the Bangladeshi Food Dataset (BFD-58), encompassing 58 distinct food items from Bangladeshi local cuisine. This contribution plays a role in the ongoing endeavor to enhance food recognition capabilities. By employing transfer learning techniques, the dataset underwent classification testing by the authors. An accuracy of 86% was achieved through single-network transfer learning, which, while noteworthy, was identified as having room for improvement. To elicit further performance enhancements, an ensemble approach involving two networks was employed, resulting in a remarkable accuracy boost of 97%. Moreover, to validate the authors' approach, comparative assessments were conducted on a benchmark dataset. The outcomes were aligned with their achievements, yielding single-network and ensemble accuracies of 86% and 94%, respectively. The dataset can be found on github
¹
.