基于深度动态联合自适应网络的图像识别方法

Image Recognition with Deep Dynamic Joint Adaptation Networks

相比传统的图像识别方法,利用深度网络可以提取到表征能力更好的特征,从而获得更好的识别效果。现实中任务提供的数据多为无标签数据或部分有标签数据,其为深度网络的学习带来了困难。而迁移学习的方法可以将从源域数据中学习到的知识迁移到目标任务的学习中,以解决有标签数据不足的问题。为了在迁移过程中减小源域和目标域间的图像数据差异,文中提出基于深度动态联合自适应网络的图像识别方法。对网络进行训练时,首先在多层网络结构中利用域间动态联合自适应方法完成针对性的数据分布自适应,然后利用熵最小化原则使学习的目标分类器穿过目标域的低密度区域,从而提高对目标域图像的识别精度。在2018年AI challenge比赛提供的24种植物病害数据集的3种迁移任务(g1->g2,s1->g2和s2->g2)中,所提方法的准确率分别达到了97.27%,94.25%和93.66%,均优于其他算法。实验结果证明,文中提出的基于深度网络并使用动态联合自适应和熵最小化原则的学习框架能够准确识别图像。

Compared with the traditional image recognition methods,the depth network can extract the features with better representational ability,so as to obtain better recognition effect.In reality,most of the data provided by tasks are unlabeled or partially labeled,which makes it difficult for deep network to learn.The knowledge learned from the source domain is used for the learning of the target domain by means of transfer learning,which can alleviate this problem.In order to overcome the image-data diffe-rence between the source domain and the target domain in the transfer process,an image recognition method based on deep dyna-mic joint adaptation networks is proposed.During the training of the transfer networks,the dynamic joint adaptation method is used to realize the data distribution adaptation in the multi-layer network structure.Then the entropy minimization principle is used for the target classifier to pass through the low-density area of the target domain.At last,the image classification and recognition are realized.The experimental results show that,with this method,the average accuracy of the three transfer tasks based on 24 kinds of plant disease provided by the 2018 AI challenge competition are 97.27%,94.25%and 93.66%,which are better than other algorithms.A large number of empirical results show that the transfer learning framework based on the deep networks,meanwhile,using dynamic joint adaptation and entropy minimization principle can recognize images accurately.