p-Faster R-CNN Algorithm for Food Detection

Eating healthily helps prevent disease, and it can be achieved by identifying the kinds and ingredients of the food to determine whether the diet is healthy. In this paper, we innovatively propose p-Faster R-CNN algorithm for healthy diet detection, which is based on Faster R-CNN with Zeiler and Fergus model (ZF-net) and Caffe framework. Before the input layer, the Gauss Pyramid is applied to form a multi-resolution pyramid of images, which expands the number and the scale of the samples. In the training stage, the multi-scale Spatial Pyramid Pooling Layer is added after the convolution layer to extract multi-scale features. To evaluate the performance of p-Faster R-CNN, we compare it with Fast R-CNN and Faster R-CNN. The experiment results demonstrated that p-Faster R-CNN increases the AP value of each kind of food by more than 2% compared with Faster R-CNN, and p-Faster R-CNN, Faster R-CNN are superior to Fast R-CNN in accuracy and speed. At last, the total dataset we established is used to construct the application of judging the healthy diet by uploading intake photos.