Bhutanese Food Recognition from Images

Despite the popularity of using deep learning techniques in food recognition, object detection, and classification in the context of Bhutanese food remains unexplored. This study presents the recognition of Bhutanese food from images using yolov8 models. The paper introduces a new Bhutanese food dataset called BHT -Food22, consisting of 22 Bhutanese dishes with 100 images in each category. The food images were collected from a few volunteers' daily food consumption and the internet. The dataset was randomly divided into 80% training and 20% validation sets. Data augmentation techniques such as scaling, flipping, translation, rotation, mosaic, and HSV color-space adjustment were applied during the training process. The dataset was trained on YOLOv8s-seg model, which uses CSPDarknet53 for feature extraction and object detection. The model was able to achieve an average mAP@50 of 0.91 for food detection and mAP@50 of 0.908 for food segmentation in the validation set.