基于卷积神经网络的食品图像识别

针对传统食品图像识别方法提取特征能力差、准确率差、运行效率差和卷积神经网络识别相似食品图像难度大等问题, 提出了一种新的食品图像识别模型ChinaFood-CNN, 以实现对食物的精准分类; 在多分类损失函数SoftmaxWithLoss的基础上, 针对食品图像类间相似性大的问题, 提出了最大类间距损失函数(MCSWithLoss), 以增大相似类之间的距离, 实现相似类的区分; 针对随机选取样本时的训练集冗余问题, 在网络模型训练时使用负样本选择算法.实验结果表明:ChinaFood-CNN模型对食品图像的识别准确率达69.2%, 分别比AlexNet、VGG16、ResNet模型提升了17.6%、16.8%和3.6%.

Food Image Recognition Based on Convolutional Neural Network

Aiming to solve the problems of the traditional food image recognition methods, such as their poor extraction ability, poor accuracy and poor running efficiency, as well as the difficulty with the convolutional neural network in identifying similar food images, a new food image recognition model ChinaFood-CNN is proposed to realize the accurate classification of food. Drawing on the research on the multi-classification loss function SoftmaxWithLoss and taking into account the intra-class similarity of food images, the maximum class-space loss function (MCSWithLoss) is proposed to increase the distance between similar classes and achieve the similar class distinction. To solve the redundancy problem of training set in random sample selection, negative sample selection is used in network model training. Experimental results show that the accuracy of the model for food images is 69.2%, which is 17.6%, 16.8% and 3.6% higher than that of AlexNet, VGG16 and ResNet respectively.