基于CBAM-InceptionV3迁移学习的食品图像分类

为提高食品图像自动识别分类的准确率，提出一种嵌入通道注意力机制和空间注意力机制的卷积块注意力模块（CBAM）的“开端”第三版（InceptionV3）分类模型。将带有图像网络（ImageNet）预训练权重参数的InceptionV3模型拆分后，在每个Inception块后嵌入CBAM模块，再重新组装成新模型，共嵌入11个CBAM模块。将此模型用于经过填充和缩放到299×299像素的Food-101食品图像数据集进行迁移学习，最高准确率达到82.01%。与原始的InceptionV3模型相比，CBAM模块能够有效提升模型的特征提取和分类能力；同时迁移学习与从头开始训练相比也可以大幅提高准确率、缩短训练时间。与其它几类主流卷积神经网络模型进行对比实验，结果表明该模型具有较高的识别准确率，可为食品图像分类识别提供有力支撑。

Food Image Classification Based on CBAM-Inception V3 Transfer Learning

To improve the accuracy of automatic recognition and classification of food images, a classification model CBAM- InceptionV3 is proposed, which embeds the Convolutional Block Attention Module. The specific method is to split the Inception V3 model with ImageNet pre-trained weight parameters into blocks, embed CBAM modules after each Inception block, and reassemble them into a new model, embedding a total of 11 CBAM modules. This new model is used for transfer learning of Food-101 food image dataset padded and scaled to 299 pixels in both length and width, with the highest accuracy of 82.01%. Compared with the original Inception V3 model, the CBAM module can effectively improve the model''s feature extraction and classification capabilities. At the same time, transfer learning can significantly improve the accuracy rate and shorten the training time compared with the training from scratch. Compared with several other mainstream convolutional neural network models, the results show that this new model has higher recognition accuracy and can provide strong support for food image classification and recognition.