Research On Large-scale Food Image Dataset Construction And Recognition

With the improved quality of life,more and more people are beginning to focus on healthy eating,this requires a deeper understanding of food,and food recognition technology can better help people in this regard.Food recognition technology has begun to be applied in real life,such as smart restaurant,smart refrigerator,and diet recommendation.However,under the background of artificial intelligence upsurge,there are some problems in food recognition technology,such as the accuracy of food recognition and the limited diversity of food categories.Therefore,improving large-scale food image recognition performance are desperately needed nowadays in the field of food recognition.Even though researchers have done a lot of works on food recognition,there is still some gap with the expectations of practical applications.The main reasons are:the scale of food image dataset is small,and there is no fully effective method for food recognition.Aiming to alleviate the above issues,this thesis has studied and discussed this topic from dataset construction and recognition methods.The main research contents and contributions are as follows:(1)This thesis constructs a large-scale food image dataset.Datasets are an important basis for various methods.A high-quality dataset can provide key guarantees for the effective verification of methods.Considering the small-scale of the existing food image datasets,this work systematically conceives and constructs a large-scale food image dataset ISIA Food-500 from constructing food categories system,collecting data,cleaning data to expanding and verifying analysis data.This dataset contains 405,776 images and 500 food categories from 52 countries.Compared with existing popular benchmark food datasets,the ISIA Food-500 is a more comprehensive food dataset with larger category coverage,larger data volume and higher diversity.(2)This thesis proposes a food recognition method based on Stacked Multi-Scale Multi-Attention Network(SMSMANet).In most existing works,people tend to use visual food image for food recognition.However,the attributes that have small inter-class differences and large intra-class differences could provide complementary information but still lack exploration such as discriminative global appearance and local details.This work proposes a Stacked Multi-Scale Multi-Attention Network to jointly learn image-oriented global and local features via combining hybrid spatial-channel attention and multi-scale strategy for food recognition.The effectiveness of the Stacked Multi-Scale Multi-Attention Network is proved by sufficient experiments.