Current Issue Cover

贾萌, 赵秦, 鲁晓锋(西安理工大学)

摘 要
目的 异质遥感图像是从不同传感器所获取的,在数据结构,分辨率及辐射特性上存在巨大差异。变化检测任务旨在通过对在不同时间获取的同一目标区域的图像进行分析来检测变化,而异质遥感图像的数据异构特性会使得变化检测过程更加困难。针对这个问题,提出了一种嵌入聚类分析的双边对抗自编码网络来实现异质遥感图像地物变化精确检测。方法 构造双边对抗卷积自编码网络对异质遥感图像进行重构和风格转换,通过结构一致性损失和对抗损失对网络训练进行约束,迫使网络将异质图像转换到公共数据域。考虑到变化区域像素对于对抗损失函数在网络优化中的不利因素,对映射到公共数据域的两对同质图像生成的伪差异图进行聚类分析,构造语义信息约束的对抗损失函数。结果 在四组典型的异质遥感图像数据集上对提出的变化检测网络性能进行了测试,分别为:Italy数据集,California数据集,Tianhe数据集以及Shuguang数据集,总体检测精度分别达到为0.9705,0.9382、0.9947以及0.9826,与现有的传统以及深度学习方法作对比,提出算法在视觉及定量分析结果上均取得了较好的检测性能。结论 针对异质遥感图像变化检测所要面临的由季节、数据异构等因素所造成的检测困难、错检率高的问题,提出的双边对抗自编码网络的无监督异质遥感图像变化检测方法,既可以实现变化检测过程完全无监督,又充分利用网络特性和语义信息,提高了变化检测性能。
Bipartite adversarial autoencoder network for unsupervised heterogeneous remote sensing image change detection

jia Meng, zhao Qin, Lu Xiaofeng(Xi’an University of Technology)

Objective Heterogeneous remote sensing images from different sensors are quite different in imaging mechanism, radiation characteristics and geometric characteristics and thus reflect the physical properties of the ground target at different levels. Therefore, there is no relationship between the observed values of the same object, which usually lead to “pseudo changes”, making the change detection task more difficult to obtain accurate change information of the observed ground objects. There has been much effort toward unsupervised heterogeneous remote sensing image change detection by designing various methods to obtain change information. However, traditional image difference operators based on the difference or ratio of radiation measurement is no longer applicable. Therefore, transferring the bitemporal images in a common space is a convenience way to calculate differences. Considering the excellent and ?exible feature learning capability, deep neural networks have been widely applied in heterogeneous image change detection tasks to effectively alleviate the influence of "pseudo changes". Moreover, fully utilizing the characteristics of Deep neural networks, it can be designed to transform heterogeneous remote sensing images into the same feature domain, and then change information can be accurately represented. Inspired by the paradigm of image translation: Heterogeneous images are transformed into a common domain with consistent feature representations, where data can be made comparisons directly. The key point of this task is learning a suitable one-to-one mapping to build a relationship between distinct appearances of images, and exclude the effect of interference factors as well. Therefore, this paper proposes a bipartite adversarial autoencoder network with clustering (BAACL) to detect changes between heterogeneous remote sensing images. Method A bipartite adversarial autoencoder network was constructed to reconstruct the bitemporal images, and more importantly to achieve the transformation of heterogeneous images to common domain. To push the process of the constructed neural network optimization to obtain good data mapping, it should be under appropriate reconstruction loss regularization. Then the bipartite convolutional autoencoders can be jointly trained to encode the input images and reconstruct them with high fidelity in output. Meanwhile, for the image-to-image translation task, additional structural consistency loss and adversarial loss terms are designed to constrain the network training process to convert heterogeneous images to the common data domain. The structural consistency loss term is designed to describe the internal structural relationships of images before and after translation. To do this, af?nity matrixes are adopted to express the structural self-similarity within an image. And the heterogeneous images can be conveniently compared in a new af?nity space. Notice that, based on the paradigm of image translation, our model can be viewed as learning two “transformation functions”, which are adopted to transform each of the heterogeneous images to the opposite data space. Furthermore, such a setup can also be seen as a special case of an “adversarial mechanism”, which is formulated by an adversarial loss to train autoencoders to match the opposite image style. Considering the disadvantage of these changed pixels to the adversarial loss term in network optimization, the pseudo-difference image generated by two pairs of homogeneous images mapped in the common data domain is analyzed by clustering in an unsupervised manner. Then the obtained semantic information is adopted to further constrain the adversarial loss term. The overall performance of BAACL network is illustrated on four sets of publicly available heterogeneous remote sensing image datasets. Five popular traditional and deep learning based heterogeneous remote sensing image change detection methods are selected and compared with this method to verify its effectiveness. Result Results obtained from Italy dataset, California dataset, Tianhe dataset and Shuguang dataset also can illustrate the performance of the proposed BAACL, with the whole overall detection accuracy up to 0.9705, 0.9382, 0.9947 and 0.9826, respectively. Meanwhile, the proposed method is superior to the five compared methods in terms of visual results of the final change map. A set of ablation experiments is designed to verify the influence of the improved adversarial loss term on network optimization performance. It can be seen that the performance of BAACL network does not change dramatically with various proportion of sample changing regions due to the semantic regularization. It verifies the effectiveness of the proposed semantic information based adversarial loss term for network optimization. Conclusion The semantic information based adversarial loss term is designed to narrow the distance of the unchanged regions of the bitemporal images for images style consistency. This is because that, for adversarial loss term without semantic regularization, the larger the proportion of the change region of the current training sample, the worse the trained network effect will be. Therefore, an accurate definition of the network constraint term will result good network optimization, and further affect the change detection performance. In view of the difficulty and high false alarm rate of heterogeneous remote sensing image change detection caused by factors such as seasons and data heterogeneity, the proposed bipartite adversarial autoencoder network for heterogeneous remote sensing image change detection can not only make full use of network characteristics and semantic information to push the image style consistency, but also realize completely unsupervised change detection process. The change detection performance can be great improved.