基于自我监督预处理的食物图像分类

Food image classification based on self supervised preprocessing

随着社交网络的快速发展,人们通常会上传、分享和记录食物图片,因此食物图像分类的应用价值也越来越大,对食品推荐、营养搭配、烹饪文化等方面都产生了积极的影响。尽管食物图像分类有着巨大的应用潜力,但从图像中识别食物仍然是一项具有挑战性的任务。为了解决食物的细粒度识别问题,本文提出了一种基于自我监督预处理的食物图像分类模型,通过自我监督的学习方式更高程度地学习食物图像特征。该模型在基于密集连接网络的食物图像分类模型DenseFood基础上搭建,采用上下文恢复的自我监督策略,将训练好的网络权重用于初始化DenseFood模型,训练微调完成分类任务。上下文恢复的自我监督策略和密集连接网络都是专注于图像特征的提取,同时结合两者,充分学习食物图像特征,来达到更好的食物图像分类精确度。为了进行性能比较,使用VIREO-172数据集对基于自我监督预处理的食物图像分类模型、未预处理的食物图像分类模型DenseFood以及基于ImageNet数据集训练预处理的DenseNet、Res Net这四个模型进行训练。实验结果表明,本文提出的食物图像分类模型优于其他策略。

With the rapid development of social networks,people usually upload,share and record food images,so the application value of food image classification is also increasing,which has a positive impact on food recommendation,nutrition collocation,cooking culture and so on.Although food image classification has great application potential,it is still a challenging task to recognize food from images.In order to solve the problem of fine-grained food recognition,this paper proposes a food image classification model based on self supervised preprocessing,which can learn food image features to a higher degree through self supervised learning.The model is based on DenseFood,a food image classification model based on dense connected network.The self-monitoring strategy of context recovery is adopted.The trained network weight is used to initialize DenseFood model,and fine-tuned trained to complete the classification task.The self-monitoring strategy of context recovery and dense connection convolution network are both focused on the extraction of image features.The research combines them to fully learn the food image features to achieve better classification accuracy of food image.In order to compare the performance,VIREO-172 data set is used to train four food image classification models:self supervised preprocessing based food image classification model,non preprocessed food image classification model densefood,and ImageNet data set based training preprocessing DenseNet and ResNet.The experimental results show that the proposed food image classification model is superior to other strategies.