Research On Fine-Grained Food Image Recognition

Fine-grained food image recognition is an important research project in the field of computer vision.With the improvement of computer hardware and the advent of intelligent life,food recognition technology will be widely used in smart restaurants,healthy diet and other fields.When the fine-grained food image is recognized by the object detection algorithm,problems such as error detection and miss detection will occur.In order to solve such problems,this thesis construct a new Chinese food dataset and propose three new food recognition methods.The main research work of this thesis are as follows:1.This thesis construct a new Chinese food dataset.This thesis determine 35 kinds of food by observing common dishes in daily diet,and use various channels to complete image collection.Since each food image contains many different Chinese dishes,there is no such dataset in the existing food dataset.Therefore,this thesis analyze the new dataset from the perspective of the number of images,the labeling box of the food,and the aspect ratio of the labeling box.It proves that the multi-food image dataset in this thesis is more advantageous.2.This thesis studies a food recognition method based on the fusion of adjacent feature information.In view of the limitations of the feature fusion method for general object detections,this thesis performs a two-step fusion operation on adjacent multiscale food features,making full use of the information contained in the food feature of different scales,improving the utilization of multi-scale feature information and increasing the detail information of the food feature.3.This thesis studies a food recognition method combining multi-receptive field attention and feature channel weighting.In view of the limitation of a single receptive field attention with less food information,This thesis propose to use multiple receptive field features to implement attention operations separately,and fuse the multiple receptive field attention.This method improve the fine-grained information of food features,and increase different levels of information in those food features.At the same time,in order to highlight the role of important channels in the food feature,this thesis again uses multi-receptive field features to implement channel weighting operations on the new attention feature,and further highlighting the importance of multi-receptive field attention.4.This thesis studies a food recognition method based on food bounding box relocation.In this thesis,starting from improving the location ability of food image,we first introduce common loss functions,and propose a general position loss function through the common loss function,and then analyze the food dataset to select a suitable loss function.At the same time,the original object detection algorithm is improved,a relocation method based on the object bounding box is proposed,and finally this method is combined with a new loss function to form the method in this thesis.In this thesis,the above work is verified on the newly constructed Chinese food dataset.The accuracy of the above three food recognition methods are 84.31%,84.60%,and 85.28%.And the experimental results show that the food recognition methods in this thesis can effectively improve the recognition accuracy and reduce the miss and error detection problems.