基于深度迁移学习的饮食图像识别研究

Research on recognition of dietary images based on deep transfer learning

卷积神经网络(CNN)应用于图像识别具有很大优势,但是需要足够深的网络和大量标签完善的数据集才能发挥其优越性。实际应用中,往往需要应对的是质量差和大小不一的数据集,且受硬件设备限制。为了提高图像识别效率和精度,提出一种基于深度卷积神经网络和迁移学习的识别算法。该算法首先对图像预处理和数据增强,后迁移大样本提取出的特征信息用于CNN特征提取,再接入微调网络对数据集再训练。实验结果显示,本文算法对饮食识别的精度和时间性能均有显著的提高,精确度最高可达98%以上,精度提升最高可达10%以上,时间性能提升幅度最高可达110%。

Convolutional neural network(CNN)has tremendous advantages when applied to image recognition,but it requires a deep network and a large number of well-labeled datasets to exploit its superiority. In practical daily life,it often needs to deal with datasets with bad quality and inconsistent size and limited by hardware devices. In order to improve the efficiency and accuracy of image recognition,a recognition algorithm based on deep convolutional neural networks and transfer learning is proposed. The image is pre-processed and augmented first,then the feature information extracted from large samples is transferred for CNN feature extraction. Finally the fine-tuned network is accessed to target datasets. The experimental results indicate that the algorithm has a significant improvement in both accuracy and time performance for diet image recognition,with the accuracy up to more than98%,precision improvement up to more than10%,and time performance improvement up to more than110%.