On the Effectiveness of Deep Transfer Learning for Bangladeshi Meat Based Curry Image Classification

Food image classification has received significant attention from researchers in recent years. A large number of methods have been shown to be working well on generic image food classification and dealing with problems like scale and appearance. However, their applicability to the classification of food images from a specific country and culture is still unknown. In this paper, we focus on Bangladeshi meat based curry image classification. We experiment with multiple transfer learning strategies to perform curry image classification with popular pretrained CNNs. Our proposed transferlearning strategies can be useful for achieving better classification performance when full fine-tuning and training from scratch of a CNN is not possible. We also proposed a new dataset for the task and performed a relevant ablation study. Our source code and dataset are available at https://github.com/MinhajurRFahad/bd-curry-image.