Research On Categorization Of Foods Based On Incremental Learning

With the enhancement of people's health consciousness,the practicability and universality of food image classification in people's daily life are increasing.Visual-based food image classification technology can not only help people quickly understand unfamiliar foods,but also can be used for nutritional composition analysis and calorie measurement,not only to help patients such as diabetes to control the sugar content consumed by food,but also to help fitness enthusiasts control the proportion of calories and various nutrients consumed.At present,most of the research on food identification is limited to western food and Japanese food,and the research in the field of Chinese food is insufficient;Chinese food contains eight major dishes,it is not realistic to obtain all dishes at one time,and the existing batch model can not meet the needs of incremental learning.Aiming at these two key tasks,this paper studies a food image classification model based on incremental learning algorithm.The main work of this paper is as follows:1.This paper proposes a new method to solve catastrophic forgetting to achieve incremental learning,weight decomposition.The idea of reducing the dependence and freezing of neurons in the two methods of dropout and fine-tune is fused.This paper designs a new method.In this method,the weight parameters in each layer are divided into two parts,one is to add the optimized parameters of the calculation diagram,the other is the non-optimized parameters without adding the calculation diagram.In the training process of deep learning,the parameters that do not add the calculation diagram still participate in the training process,but because they are not in the calculation diagram,the non-optimized parameters can not participate in the back propagation,and the optimization parameters of the optimizer can not be received.this method by controlling the optimization direction of some parameters can increase the plasticity in stable plasticity dilemma,also control the stability of the model and mitigate the effects of catastrophic forgetting.2.This paper combines the characteristics of feature reuse in the Dense Net based on the Res Net of residual structure,and combines multi-scale features to realize a multi-scale fine-grained image classification method,which improves the use of fine-grained information by the model,and makes up for the disadvantage of information loss after the increase of network layers.We rearrange the network structure to improve the balance between each feature layer,mainly to improve the extraction of the underlying location information.The features of each layer are fused by the method of semantic embedding.First,the bottom layer and middle layer are fused,the fusion results are fed into the next layer to extract the high layer features and then weighted fusion is carried out.Finally,the features of each layerare fully fused and then fed into the classification layer for classification operation.A small data set was created to verify the effectiveness of the proposed method,using the lab to crawl pictures of some dishes from various recipe websites in 80 categories,160,000 pictures CF80.After Chinese Food Net pre-training,then transfer to the CF80 for learning fine-tuning,and the training test of incremental food classification model.First,we use the basic MNIST data set to test the effectiveness of the algorithm,and the test shows that our algorithm has the effect of incremental learning,which is applied in the final chapter in the classification of dishes.The experimental results show that our method has the effect of fine-grained incremental learning on food classification and has the effect of slowing down catastrophic forgetting.