Deep Learning Based Food Detection System Research

With the rapidly development of artificial intelligence and the increasing awareness of people’ health,the research about food detection has been attracting much attention.Extended applications of food recognition such as health management systems and automatic restaurant billing systems can improve dining happiness and social productivity.Traditional image processing methods are difficult to quickly solve data with many classes and large sample differences.In recent years,deep learning has been widely used in computer vision tasks,so this paper studies food recognition and detection models based on deep learning methods,and the main research work is as follows:Firstly,we designed IRASNet(Inverse Residual Attention Channel Shuffle Convolutional Network),a lightweight food recognition model,to address the problem that Chinese foods are various and highly similar,and it is difficult to recognize foods due to lighting and occlusion during the filming process,etc.We used deepthwisesparable convolution to reduce the amount of operations in the convolution process,and introduced SE-Net and Channel Shuffle in the inverse-residual structure to improve the feature extraction ability of the network.The experimental results show that using random erasure to simulate the effect of occlusion during training can improve the model robustness,meanwhile,IRASNet achieves 81.88% accuracy on Chinese Food Net,and 91.26% accuracy on Food101.Secondly,for the problems of large size of SSD object detection model,slow detection speed and difficult to achieve good results on platforms with weak computational power,a multi-scale lightweight food detection model IRAS＿SSDLite is designed,and a food detection visualization system is constructed based on this model.SSD backbone network uses lightweight food recognition IRASNet network to reduce the amount of convolutional operations in the feature extraction part;the Euclidean distance between bounding-box is used as a penalty term to improve the IOU calculation,and the regression loss function is adjusted accordingly.The experimental results show that the detection accuracy of IRAS＿SSDLite is 97.3% in the self-built 20-class dataset Chinese Food,which is comparable to SSD;it can detect 105 images per second,and the frame rate is 129% higher than SSD model;the model size is 11.5MB,and the size is 89% less than SSD.