A Chinese Dish Detector with Modified YOLO v3

Abstract

In the field of food recognition, considering the difficulty in feature extraction caused by the diversity of Chinese dish, this paper proposed a Chinese dish recognition method SeDC-YOLO based on Squeeze-and- Excitation (SE) and Deformable Convolution (DC). This method optimized the feature extraction network Darknet53. Specifically, the SE attention mechanism was adopted to model the correlation between the different channels of the feature map as well as strengthen important features. Meanwhile DC was taken to solve the problem that the network is difficult to adapt to geometric deformation caused by regular sampling in standard convolution. Furthermore, the Gradient Harmonizing Mechanism (GHM) was introduced to solve the problem of unbalanced classification samples. Aiming at the local recognition problem in dish recognition, a new Non-Maximum Suppression (NMS) strategy was proposed. The experimental results of 37 types of Chinese dishes showed that the method proposed in this paper has a higher recognition accuracy.