Machine-vision-based Bangladeshi street food recognition

Abstract

Bangladeshi street food undoubtedly forms an important aspect of culinary culture, however, its automatic identification has been largely unstudied because of the lack of appropriate datasets and visual similarities among food categories that already exist. This paper offers a framework for recognizing Bangladeshi street food based on machine vision and deep learning. The contribution of the study lies in the fact that it offers a custom dataset of 4,000 original images and then augmented it to 16,000 images. Five pre-trained CNN models (InceptionV3, DenseNet121, MobileNetV2, NASNetMobile, Xception) were evaluated using the standard metrics. The baseline model of MobileNetV2 demonstrated the accuracy of 93.75%. Further research was based on two-stage transfer learning. The model obtained 99% accuracy, precision, recall, F1-score, which is much more than that obtained by the MobileNetV2 model. Grad-CAM and LIME technologies were used for better interpretability of the learning model.