Research On Chinese Food Image Recognition And Application Based On Optimized YOLOv4

To deploy the Chinese food image detection model on mobile devices,this paper proposes a Chinese food image detection model based on optimized YOLOv4 and proves its effectiveness through experiments.After that,based on the proposed model,a Chinese food image detection system is designed and implemented.The main research contents and work of this paper are as follows:1.To address the problems that traditional target detection models take up more resources and have lower accuracy in lightweight detection models,this paper first improves the YOLOv4 model from the perspective of lightweighting,using the lightweight neural network Mobile Net V2 to replace the original YOLOv4 backbone feature extraction network CSPDark Net53 on the one hand,and using the 5×5 depth separable convolution to replace the PANet and the 3×3 normal convolution in the YOLO-Head structure,thus effectively reducing the size of the model.Second,to compensate for the performance loss caused by the lightweight of the model,the ECBAM(enhanced convolutional block attention module)attention mechanism module is added after the PANet structure to enhance the feature extraction capability of the model.Finally,a Chinese food image detection model(CFIDM)based on the optimized YOLOv4 is proposed.It is experimentally demonstrated that the model reduces the volume to about 20% and improves the speed to about 1.6 times compared with the original YOLOv4,while basically maintaining the original performance.2.To solve the problem of small size of the new dataset,the training method based on migration learning fine-tuning is used to effectively improve the detection accuracy of the model;the K-Means clustering algorithm is used to re-cluster the prior frame on the Chinese food image dataset,which solves the problem of slow convergence of the YOLOv4 preset prior frame on the new dataset;finally,to solve the problems of missed detection and repetition of the model in the prediction process,a non-maximum suppression algorithm based on Soft-NMS(soft non-maximum suppression)optimization is given to solve the problems of missed detection and repeated detection in the model prediction process.It is proved that the above method can effectively improve the performance of the CFIDM model.3.Based on the above research results,we designed and implemented a Chinese food image detection system with the functions of food image recognition,data query,and model update by combining mobile technology and frameworks such as Flask and Vue,and verified the usability of CFIDM model.Main contributions: A Chinese food image detection model based on optimized YOLOv4 is proposed,which reduces the size of the model while basically maintaining the performance of the original model;a dual-threshold non-maximum suppression algorithm based on Soft-NMS optimization is given;a Chinese food image detection system is designed and implemented.