A beautiful image or not: a comparative study on classical machine learning and deep learning

With the development of web services and Apps on the Internet, food images are emerging into our life. Consumers from yelp or the dianping service upload a lot of food pictures every day. The images usually express the users' feelings and are shared among the social network. There have been different researches on the images. However, there is few research on how to evaluate the food image is beautiful or not. Therefore, we came up with an idea to classify food pictures by their appearance, which is meaningful in multiple applications, especially picking beautiful pictures to help businesses attract customers. In order to realize this idea, we collected 1067 food images through web crawling and questionnaires. Each image has a unique label: beautiful or not beautiful. Machine learning methods are used in this paper to model the data. CNN models in deep learning: VGGNet, AlexNet, and ResNet can get good results, e.g., ResNet can achieve the accuracy of 95.83%. However, with a good feature engineering job, the classifiers, which are random forest and support vector machine can reach a better accuracy of 99.63%. The experimental results indicate feature engineering is a vital issue in the food image evaluation problem, which lacks of labeled data.