Research On Food Intake Action Recognition And Food Detection Algorithm Based On Artificial Intelligence

Relevant surveys and studies have shown that more than 2.4 billion people in the world suffer from the double burden of malnutrition,and effective dietary intervention can induce people to focus on healthy diet behaviors and nutritional balances,it is a means to mitigate the negative effects of the double burden of malnutrition.Therefore,the following two aspects were highlighted in this paper:(1)A food intake identity recognition method that can apply to record the diet behavior information based on recurrent neural network was proposed.First of all,three-axis accelerometer was applied to develop a smart spoon that can record food intake action in real time;secondly,model for food intake action and identity recognition was built based on longshort time memory neural networks;finally,the performance of smart spoon and models were validated by many experiments.The results showed that the accuracy of food intake action and identity recognition models can reach 94.58% and 93.97% respectively,it indicated that method mentioned above can record diet actions accurately indeed.(2)On the other hand,this paper first collected and produced a Chinese food dataset that covers 66 food categories such as hot pot,porridge and special snacks and includes about more than tens of thousands of instances;then a food object detection model with related improvements was built based on YOLOv5,especially for the situation of unbalanced distribution on the dataset(or Long-tailed distribution)in objectness and classification branches,loss functions that existed or was proposed in this paper based on gradient reweighting were integrated into model in order to improve the performance;finally,the results of the models and effects of hyperparameters were detailed tested and verified by many experiments,the results has shown that m AP.5?m AP.5:.95 can reach 90.03% and67.59% respectively.It shows that it is feasible to estimate the nutritional composition of daily diet with the help of the food detection model constructed in this paper.In addition,a graphical user interface was designed to integrate the above models and includes correction function of false positive detections.