MaskMix:用于变化检测的掩码混合数据增强方法

MaskMix:mask mixing augmentation method for change detection

数据增强是提升变化检测模型泛化能力的一种主要方法。尽管现有的数据增强方法在图像分类、目标检测中取得了较好的效果,但忽略了多个时间序列图像之间的差异和变化目标的多样性。为了较好地保留变化区域并且增加复杂的背景信息,基于变化区域掩码,提出一种适用于变化检测的数据增强方法:MaskMix。首先,将当前图像对的变化区域粘贴到一个图像对上,得到具有新的背景和新的变化的变化图像对。其次,采用多路径加权融合策略进一步增强变化图像对。在每条路径上,从图像处理集合中随机选取一种经典的图像处理方法进一步处理变化图像对,然后使用Dirichlet分布产生的K维权重将K条路径处理后的图像对进行融合。最后,通过跳跃连接将处理前的图像对和处理后的图像对按Beta分布产生权重,进行更深层次的混合。实验结果表明,提出的MaskMix在BCD和LEVIR-CD两个数据集上,有效地提升了变化检测方法ADCDNet、BIT、ChangeFormer、SNUNet和DSAMNet的泛化性能。与现有的图像增强方法MixUp、AugMix、MUM和CropMix相比,MaskMix能有效增加变化图像的复杂性和多样性,提升现有变化检测方法的泛化性能。

Data augmentation is a key technique to improve the generalizability of change detection models.Although the exis-ting data augmentation methods achieve promising performance in image classification and object detection,they ignore the differences among the time-series images and the diversities of the changed objects.In order to preserve the change region and increase the information of the complex background,this paper proposed a novel data augmentation method for change detection based on change region mask,called MaskMix.Firstly,the change regions of the current image pair were pasted on an image pair to generate a new change image pair having new background and new changes.Secondly,MaskMix further augmented the image pair by multi-path weighted fusion strategy.It selected a classical image processing method randomly from an image processing set for each path to conduct further augmentation.And then the processed image pairs from K paths were fused using a K-dimensional weight generated by Dirichlet distribution.Finally,the pre-processed image pair and the post-processed image pair were fused by the weight generated by the Beta distribution through the skip connection.Experiments conducted on two publicly available datasets,e.g.,BCD(build change detection)and LEVIR-CD(LEVIR building change detection dataset),demonstrate that MaskMix significantly improves the generalizability of change detectors,e.g.,ADCDNet,BIT,ChangeFormer,SNUNet,and DSAMNet.Moreover,compared with the existing image augmentation methods,such as MixUp,AugMix,MUM,and CropMix,MaskMix effectively increases the complexity and diversity of change images,enhancing the generalizability of existing change detection methods.