Food Classification And Detection Based On Convolutional Neural Network

With the development of China's economy and the improvement of human living standards,obesity,overweight and diabetes are also increasing in China.These diseases are largely caused by unhealthy diet.At present,the hospital or institute still uses a diet log or questionnaire to track patients' diet.Therefore methods are both troublesome and inaccurate.In recent years,smart wearable devices have developed rapidly,so it's possible to automatically identify the food image by using a wearable camera to record a day's diet.Therefore,it is of great practical significance to recognize the food image automatically and classify it,improve its detection precision and reduce the size of the model.Traditional image detection algorithm has many researches in food detection and classification,but for a variety of food and complex external environment,the traditional image detection algorithm is very weak,and the general accuracy rate is too low.Deep learning,as a rapidly developing technology in the past few years,shows the ability to dominate many image detection problems.The convolution neural network is used to identify the food images,and the accuracy is very high.But the convolution neural network has a large number of parameters and complex computation.At present,algorithms based on convolution neural networks tend to run only on large servers or machines equipped with GPU.Aiming at its shortcomings,the traditional food and image convolution neural network are analyzed and summarized to simplify model of convolutional neural network is designed and apply to the detection and classification of food images,reducing the size of the model in this paper.Firstly,we study the traditional food image detection algorithms and introduce and the principles of the algorithm in detail,their shortcomings are pointed out.Then we introduce the convolution neural network,and the mathematical principles of network structure,and the network training are introduced in detail.The image detection algorithm based on deep learning is introduced in detail.Then we deeply analyze the convolution operation,and propose deep separation convolution,simplify convolution operation,propose global mean pooling to reduce network parameters,and propose a food image classification algorithm.In this paper,a food image recognition algorithm is proposed by combining the depth separation convolution with the image detection algorithm.Secondly,the working principle and construction steps of the deep learning framework(Tensor Flow)are introduced in detail.And then the self-captured food images and the images extracted from Image Net are made into the data sets that meet the requirements of this paper.The network model of this algorithm is built on the Tensor Flow framework and its core algorithm is introduced.Then the Momentum optimization algorithm is used to detect the food detection algorithm Training.Adam optimization method is used for food classification algorithm training.So we get food detection network and food classification network.And at last the model was transplanted to the Android platform.Finally,we use the test sets to test the food detection algorithm,food classification algorithm and several other algorithms respectively,summarize the three aspects of recognition accuracy,real-time and size,and the accuracy and running time of the model are analyzed on the mobile phone.Compared with the traditional food detection algorithm and deep learning food classification algorithm,the algorithm has high accuracy,good real-time,and very small model,and can successfully carry out a complete transplant to the portable device.