Research On Automatic Food Recognition Algorithm Based On Deep Learning

The rapid development of science and technology has greatly improved people's daily life.As an important application of computer vision in real scenes,automatic food recognition is one of the current research hotspots.Automatic food identification can not only be used for automatic price calculation,increase the settlement window,reduce human resources loss,effectively reduce customers' waiting time in line,but also can be used for customer's eating preference analysis.Subsequent analysis of eating habits in this area can be conducted through big data,which can guide inventory management.At the same time,it can also analyze the nutrition of the customers themselves,helping customers to eat a reasonable diet and avoid the occurrence of metabolic diseases.There are many deficiencies in traditional image processing and machine learning methods in automatic food recognition algorithms.Most of the studies are based on food classification tasks.In practical scenarios,such methods are susceptible to factors such as light intensity,noise interference,and position and direction.With the development of deep learning,convolutional neural networks have achieved great success in areas such as image recognition,object detection,and semantic segmentation.Food recognition technology is also gradually being researched around convolutional neural networks.However,during the actual implementation of food recognition,it was found that there are problems such as high sample tag acquisition costs,long food update models,long update cycles,and slow recognition speed.This paper is devoted to solving the above problems,researching bowl detection and food recognition in the actual canteen application billing scenario.The main work includes the following aspects:For the first time,construct and open source the bowl dataset Bowl-10 and bowl dataset Bowl-95 in real scenarios at home and abroad.Datasets are used for bowlobject detection research.The Bowl-10 dataset contains 31111 samples,and the Bowl-95 dataset contains 54860 samples,which containing 10 and 95 types of bowls,respectively.In addition,the Bowl-10 dataset additionally divides the actual test samples used,including various background interference factors,such as mobile phones,chopsticks,and payment cards,which is conducive to testing the ability of the model to generalize.Construct and open source the first domestic and foreign large-scale object detection food dataset CNFood-252 and the Few-Shot learning dataset FewFood-50.The CNFood-252 dataset contains 32190 samples,totaling 252 categories of common Chinese foods.The FewFood-50 dataset contains 50 types of food,each type contains a small number of samples,and is mainly used to study the problem of food recognition with a small number of samples.A cross-connect Faster R-CNN bowl detection model combining low-level features and high-level features is proposed.Using visualization to analyze in detail the general object detection model Faster R-CNN for the detection principle and process of bowl object,and propose a cross-connect Faster R-CNN model based on the characteristics of the bowl and the visualization results.On the basis of the Faster R-CNN model,a cross-connect layer and a feature fusion layer are added.The core is to reuse and fuse the shallow and deep features extracted by CNN to make up for some of the information lost in the CNN pooling process and strengthen the feature description of bowls,Improve the detection accuracy of the model while ensuring that the detection speed is basically unchanged.Experiments on the proposed cross-connect Faster R-CNN bowl detection model in the Bowl-10 and Bowl-95 datasets and comparison with Faster R-CNN,the results prove that the proposed cross-linked structure can increase the accuracy of dish detection when the increased detection time is within an acceptable range.Decoupling food recognition problems into food positioning and food classification problems and proposing a small Few-Shot learning model YOLO-Siam.The Tiny-YOLO model has extremely fast detection speed and good localization ability.Therefore,Tiny-YOLO is trained on the Bowl-95 dataset for food positioning.Food classification can choose a suitable classification model according to actual needs.on the CNFood-252 dataset,the positioning reclassification method proposed in this paper improves the accuracy of food recognition.The YOLO-Siam model combines Tiny-YOLO and the twin network,and attempts to solve the problem of large number of samples in food recognition through Few-Shot learning.Although the YOLO-Siam model did not achieve high accuracy in the FewFood-50 dataset,it pointed out a feasible direction for subsequent research.