Food image recognition using deep convolutional network with pre-training and fine-tuning

In this paper, we examined the effectiveness of deep convolutional neural network (DCNN) for food photo recognition task. Food recognition is a kind of fine-grained visual recognition which is relatively harder problem than conventional image recognition. To tackle this problem, we sought the best combination of DCNN-related techniques such as pre-training with the large-scale ImageNet data, fine-tuning and activation features extracted from the pre-trained DCNN. From the experiments, we concluded the fine-tuned DCNN which was pre-trained with 2000 categories in the ImageNet including 1000 food-related categories was the best method, which achieved 78.77% as the top-1 accuracy for UEC-FOOD100 and 67.57% for UEC-FOOD256, both of which were the best results so far. In addition, we applied the food classifier employing the best combination of the DCNN techniques to Twitter photo data. We have achieved the great improvements on food photo mining in terms of both the number of food photos and accuracy. In addition to its high classification accuracy, we found that DCNN was very suitable for large-scale image data, since it takes only 0.03 seconds to classify one food photo with GPU.