Research On Food Materials Image Recognition Algorithm Based On Deep Transfer Learning

According to the "2018 China Catering Industry Annual Report",there are nearly 3.9 trillion catering market each year,of which 1.5 trillion is the purchase price of food materials.At present,we have developed intelligent electronic scales for the automatic procurement of food materials,which can automatically identify and weigh the food materials.However,due to the complex and diverse receiving environment of the restaurant,the quality of the image of the food materials is susceptible,and the distribution of the food materials data acquired by each restaurant is also very different.Therefore,there is a transition from the laboratory to the real environment,that is,the training algorithm using the laboratory-based training data cannot be well applied to the actual environment.On the other hand,since the current dataset cannot meet the application of the electronic scale system,the Meal-300 data set is constructed,which contains 300 kinds of ingredients,about 53,374 images,but the number of each category is between 10 and 453,Between,which leads to the problem of unbalanced food materials data,that is,due to the large scale of the species,the training data of different categories is seriously unbalanced,which will lead to over-fitting problems in the category of food materials with less training data.In order to solve the above two problems,the transition from the laboratory to the real environment and the imbalance of food data.This paper proposes a deep transfer learning method for food materials image recognition in complex environments.This method is based on domain confusion and a priori tree CNN model(TAN).Compared with traditional CNN,this model has two innovations: domain confusion across the fully connected layer and a priori tree for classification.Once the source training data(lab dataset)and target training data(real-world data)are delivered to the deep learning model,CNN layers process the data as in the traditional manner,then,domain confusion is conducted through those fully-connected layers to distil the invariant features of crossing domains,and finally,the priori knowledge tree is enforced for the data classifications in order to overcome the over-fitting problem brought by the imbalanced training data.The TAN model is compared with the existing domain confusion model,and the accuracy and feasibility of the model are verified.The experimental results show that the image recognition of the food materials in the complex environment reaches an average accuracy of 58.75%.In actual electronic scale system applications,the orders purchased each day for each restaurant are very similar.Therefore,the food materials image recognition task of each restaurant can be regarded as an independent target domain.In this field,the food materials are directly identified by the weight of the food materials.This comparative weight feature can be incorporated into the target domain classification without regard to domain differences.Therefore,the TAN model is optimized,and the TAN Scales model is proposed for the practical application of restaurant food identification.The Order-31 dataset is proposed for experimental research.By comparing the TAN Scales model and the TAN model,the order information can be obtained.The TAN Scales model can further improve the accuracy of food materials image recognition.The experimental results show that in the specific application,combined with the order information,the average accuracy of the recognition of food materials is as high as 72.45%.Finally,based on the TAN Scales model,the food materials image recognition system was designed and implemented for the purchase of food materials in restaurants.The system can quickly and accurately identify the category of food materials.