Research On Food Image Classification Based On Deep Learning

Food image classification plays an important role in intelligent health management.With the continuous improvement of software and hardware and the development of the theory about artificial intelligent,food image classification will be more practical and universality on auxiliary life and social entertainment.As a subtopic of fine grained image classification,leverage details of image and locate discriminative region to guide food image classification has obtained a general concern and research.In this thesis,three novel food image classification algorithms are proposed for the common 90 categories food image.The main work of this thesis are as follows:1.In this thesis,a multi-scale method based algorithm for food image classification is studied,including two aspects: multi-scale input image and multi-scale convolutional neural network feature fusion.In terms of multi-scale input image,this thesis utilizes two kinds of image scale as input information to train the CNN classifier,by the time,two different input scales have their own network and don't share the network parameters.In terms of the multi-scale convolutional neural network feature fusion,firstly,the low-level,middle-level feature maps are weighted fusion.Secondly middle-level,high-level feature maps are weighted fusion also.Finally use the fusion feature to classification.The multi-scale based algorithms can make full use of the details of images,and make up the defects of details dropout with the deepen of networks.2.This thesis studies a discriminative region guided food image classification algorithm.Cause the common backgrounds of food images are desks or tablecloth,in order to locate the region of food,we first segment the tableware and eliminate the noise of background.Secondly,the saliency map is extracted of the image.At the same time,comparing the Interest-over-Union between food region and saliency,and calculate the final discriminative region of food images.Finally,extract the feature of discriminative region for classification.3.An attention model based food image classification algorithms is studied in this thesis.We design an attention extract model for food image,and apply it to the CNN network of this thesis for obtaining the attention feature map.Then the extracted attention feature map is utilized to improve the origin feature map.Moreover,we cascade the transformed CNN feature map and the full connect feature map to make up the loss image spatial information in full connect layer.For the sake of training the CNN model and verification of the algorithms in this thesis,we construct a food image dataset CF90 through crawling the menu website pictures.For more effective results,the algorithms are not independe nt,every new algorithm is relying on the studied achievement and making a difference.The experimental results show that the algorithms can gradually improve the classification accuracy of the food images.