基于熵增强的无监督域适应遥感图像语义分割

为了实现利用有标注源数据获得在无标注目标数据上可用的遥感图像语义分割模型,提出了一种基于熵增强的域适应端到端语义分割方法.首先,为了充分利用遥感图像多尺度信息并且减少域之间传感器分辨率带来的域偏移,采用空洞空间金字塔池化模块作为分类器;其次,为了使无标注的目标域类别正确对应,使用了两个分类器进行协同训练;将像素点预测值的信息熵当做分类置信度的度量,将其作为对抗损失的权重,从而使训练能专注于难分类的像素,降低域偏移.在ISPRS(WGII/4)2D数据集上进行实验,所提方法相对于直接使用分割模型和使用传统对抗方法,mIoU分别提高了18％和12％.实验结果表明,所提方法在遥感图像域适应语义分割表现上优于直接使用分割模型或使用传统对抗域适应分割方法.

Entropy enhanced unsupervised domain adaptive remote sensing image semantic segmentation

In order to obtain a remote sensing image semantic segmentation model that could be used on unlabeled target data by using annotated source data, this paper proposed an end-to-end entropy-enhanced domain adaptive semantic segmentation method. First, in order to make full use of the multi-scale information of remote sensing images and reduce the domain shift caused by sensor resolution between domains, the method used the atrous spatial pyramid pooling module as the classifier. Second, in order to correctly correspond to the unlabeled target domain categories, it used two classifiers for co-training. Then, it used the information entropy of the predicted value of the pixel as the weight of adversarial loss which was a measure of the confidence of the classification, so that the training could focus on the pixels that were difficult to classify and reduce the domain shift. Experiments on the ISPRS(WGII/4) 2D dataset, the mIoU of the proposed method is 18% and 12% higher than that of the direct use of segmentation model and the traditional adversarial method respectively. Experimental results show that the proposed method performs better than the direct use of segmentation model and traditional adversarial domain adaptive segmentation methods in remote sensing image domain adaptive semantic segmentation.