Research And Application Of Ingredients Recognition Method Based On Deep Learning

Ingredients are the basic material conditions to ensure people’s survival and development.Obtaining timely and accurate information about ingredients is one of the important directions of the development of intelligent identification technology.Intelligent identification is of wide application in various scenarios such as automatic picking of ingredients,intelligent trading,and refrigerator storage.The traditional identification methods are too costly and inefficient to complete the complex ingredients task.The recognition methods based on deep learning have the advantages of high accuracy and speed in the thesis,and the models have good flexibility and expansibility.The main research contents are as follows:In the aspect of basic theory and data sets,firstly,the overall structure,loss function and evaluation index of convolutional neural network are studied,and then the encoder,self-attention mechanism and decoder of Transformer network are studied,and the network structure characteristics of the two types of models are compared.Finally,the data sets for ingredients classification and ingredients detection are constructed,and the data expansion methods of single image deformation and multiple images mixing are studied to complete the expansion of the data sets.In the research of ingredients classification model,firstly,the characteristics of three classical classification network models are analyzed.Then,aiming at the accuracy and convergence performance of the classical models,the ingredients classification model ICNet based on convolution neural network and combined with Transformer model is proposed.Comparative experiments are conducted on the ingredients classification dataset,and the results show that the accuracy of the ICNet model reaches 98.02%,with a model parameter size of 15 M,and the convergence process is relatively stable.At the same time,different normalization methods and activation functions are introduced,and the results show that the ingredients classification model incorporating BN and Gelu has the best recognition effect,with the classification accuracy of 99.26%.Finally,further analysis is conducted on the effect of the positions of BN and Gelu on the performance of the model,and the results show that the performance of the model of adding transformation layers at positions I and II of the bottlenecks is better,with the classification accuracy of 99.14%.In the research of ingredients detection model,based on the research of classification models,firstly,the characteristics of three classical target detection models are analyzed.Then,an improved model based on YOLOv7 is proposed,and the backbone network introduces the self attention module B-MHSA as the basic model IDNet.Comparative experiments are conducted on the ingredients detection dataset,and the results show that the m AP value of the IDNet reaches 94.35%,with an inference time of 27.53 ms,and the parameter size is slightly increased.Then fusion experiments are carried out on feature fusion processing network and loss function,and the results show that the IDNet with both Bi FPN and Wi OU has the best comprehensive performence,with the m AP value of96.24%,and an infernce time of only 29.86 ms.Finally,the compression methods based on pruning are studied,the results show that when the compression ratio is below 0.3,the m AP values of the model decrease by an average of only 4.21%.In the application of ingredients detection model,firstly,the characteristics of different model deployment methods are analyzed,and a demonstration of the mobile end of detection effect of the model are completed.Then,the collection and transmission of ingredients images are introduced,and an ingredients information management App is developed for use with the ingredients detection algorithm,and the software testing are completed.