遥感图像小样本舰船识别跨域迁移学习算法

Cross-domain transfer learning algorithm for few-shot ship recognition in remote-sensing images

跨域迁移学习旨在利用现有公开数据集,突破源域和目标域样本类别空间须一致的约束,提升目标域样本的识别精度。针对现有跨域迁移学习算法应用于遥感图像小样本舰船目标识别时存在的迁移类别受限和负迁移问题,本文提出一种基于源域样本相关性排序的跨域迁移学习算法:首先将目标域样本逆向加入源域分类任务中,根据加入前后各类别源域样本的识别精度变化情况,对源域样本进行相关性排序,将其划分为强/弱/负相关样本;然后采取自监督联合学习策略,在目标域分类网络中引入自监督角度预测辅助分支,筛选出的强相关源域样本仅参与该辅助分支的训练,不改变目标域主分类网络的结构。算法通过相关性排序去除了弱/负相关源域样本,有效避免了负迁移;引入自监督角度预测辅助分支,在保持主分类网络结构完整性的同时,充分利用了强相关源域样本的有效信息,学习到更具泛化能力的目标特征。实验结果显示:在遥感舰船小样本目标数据集上,提出的算法优于跨域迁移学习中广泛使用的Fine-tune (微调)算法;与仅使用主分类网络的目标域识别算法相比,遥感舰船目标识别精度提升了17.59%。

Cross-domain transfer learning aims to utilize public datasets as source data to improve the recognition accuracy of target data,breaking through the limitation that the category space between source data and target data must be consistent.For the few-shot remote-sensing ship recognition task,existing cross-domain transfer learning algorithms have the disadvantages of transfer category restriction and negative transfer effect.Therefore,a cross-domain transfer learning algorithm based on source data correlation sorting was proposed to solve the above problems.First,the target data were added reversely into the source domain recognition task.According to the variation of the source data recognition accuracy before and after the target data were added,various source data were classified into strong/weak/negative correlation samples,and only the strong correlation samples would be selected.Then,the self-supervised joint learning strategy was adopted to introduce the auxiliary self-supervised angle prediction branch into the classification network in the target domain.The selected strong correlation source samples were added but only into the training of the self-supervised branch,which avoided changing the main classification network structure.Randomly selecting 65 category samples as the source data from miniImageNet and conducting comparative experiments on few-shot ship targets in remote-sensing images yields the following results:1)When Resnet18 is chosen as the classification network,the performance of the proposed algorithm is better than that of the Fine-tune algorithm,which is widely used in cross-domain transfer learning.Moreover,compared with the recognition algorithm,which only uses the main classification network,the proposed algorithm improves the recognition accuracy of target data from 78.89%to 96.48%.2)Using different networks to sort correlations for the source data,the selected strong correlation source samples are not exactly the same and their degree of category coincidence is close to 60%.However,they are all helpful to the classification task of the target domain.At the end of this paper,through visualizing the extracted target features,it is verified that the target features extracted by using the proposed algorithm are more abundant and have higher generalization ability.The proposed algorithm has two main advantages.First,the weak/negative correlation source samples are eliminated by correlation sorting,which can avoid the occurrence of negative transfer effect.Second,by introducing the self-supervised angle prediction branch,the information of the strong correlation source samples is fully utilized and the features with more generalization ability are extracted while maintaining the structural integrity of the main classification network.