Research On Food Image Classification Based On Convolutional Neural Network

Obesity is one of the most common diseases caused by nutritional disorders in contemporary society.The prevalence of obesity can be controlled by accurately identifying and giving the fat content and calories in foods.This requires the ability to classify food effectively,but manual food classification is no longer appropriate in today's fast-paced society.The large number of food categories,and the high similarity between many of them,makes classification difficult and can be categorized as a fine-grained classification problem.In fine-grained image classification,the computational cost and expression dimension of bilinear convolutional neural network models are high,and it would be of great practical importance if the computational cost of bilinear structures can be reduced while also improving(or not)the classification accuracy.In traditional image recognition algorithms,the algorithms have poor performance,low accuracy,and usually do not enable an end-to-end learning process.In recent years,the technique of convolutional neural networks in deep learning has a good performance in many image recognition fields.In the classical convolutional neural network model,its model itself is larger and has more parameters,while the bilinear structural model has a low feature expression ability and a large expression dimension when performing fine-grained image classification.To address its shortcomings,this paper improves the method from two different perspectives on the basis of bilinear convolutional neural networks,so as to overcome the limitations of the original bilinear structural model.(1)Most algorithms treat fine-grained image features as a whole and ignore the spatial relationships between the target parts(local).Spatial relationships between target parts are useful in tasks such as image retrieval,facial classification,and image classification.This paper therefore proposes to exploit the spatial relationships between image features to obtain connections between them.Spatial relationships can enrich the understanding of images and provide additional basis for image classification.(2)Similar to the second-order statistical features in bilinear structures,spatial relationships are often used only to express autocorrelation information.Therefore,this paper proposes to classify food images based on second-order statistical features across different layers,and proposes three cross-layer feature fusion methods to obtain better performance.Finally,experimental evaluation is carried out on the model size,number of parameters,expression dimension and classification performance.