A Method of Constructing a Food Classification Image Dataset by Cleansing Web-Crawling Data

抄録

We propose the construction of image datasets via data cleansing for food recognition using a convolutional neural network (CNN). A dataset was constructed by collecting food images and classes from web crawling sites that post cooking recipes. The collected images included images that cannot be effectively learned by the CNN. Examples include images of foods that look extremely similar to other foods, or images with mismatched foods and classes. Here, these images were termed “content and description discrepancy images.” The number of images was reduced using two criteria based on the food recognition results obtained using CNNs. The first criterion was a threshold for the difference in the estimated probabilities, and the second was whether the estimated class and food class matched. These criteria were applied using multiple classifiers. Based on the results, the dataset size was reduced and a new image dataset was constructed. A CNN was trained on the constructed image dataset, and the food recognition accuracy was calculated and compared using a test dataset. The results showed that the accuracy using the dataset constructed using the proposed method was 7.4% higher than that of the case using web crawling. This study demonstrates that the proposed method can efficiently construct a food image dataset, demonstrating the data-cleansing effect of the two selected criteria.