SAR影像变化检测的前景特征流形排序法

Foreground feature manifold ranking method for SAR image change detection

SAR影像变化检测的差异图分析法存在的两个问题:①连通区域内的部分变化区域易被误判为未变化区域;②中心先验假设并不适用于检测位于SAR影像边界的变化区域。本文针对以上两个问题设计了一种超像素分割和前景特征流行排序(manifold ranking,MR)的SAR影像变化检测方法(MRSFCD)。首先,通过单像素和邻域对数比算子进行加权融合构造差异图,可以保持变化区域内部的一致性并抑制噪声干扰。其次,对差异图进行超像素分割。然后,改进超像素的无向图连接方式,不对边界四周的超像素进行连接,利用超像素分割结果和灰度信息进行三次邻接。最后,将基于前景特征流行排序后得到的显著性图与单像素对数差异图进行点乘,对其进行阈值分割得到最终的二值变化图。本文通过采用3组双时相影像进行试验。结果表明,相较于其他变化检测算法,本文方法有效地提高了变化检测的精度。

There are two problems with the difference image analysis for the current SAR image change detection methods.Some of the changed areas in the connected area are easily misclassified as unchanged areas,and the central prior assumption cannot be well applied to detecting the changed regions located at the boundary of the SAR image.In order to avoid the above limitations,a method of manifold ranking based on superpixel segmentation and foreground features for change detection(MRSFCD)was designed.Firstly,the difference image was constructed by weighted fusion of single pixel and neighborhood logarithmic ratio operator,which can maintain consistency within the change areas and suppress noise interference.The difference image was then segmented by the superpixel model.After that,the improved undirected graph connection method of superpixels was proposed.The main idea is that superpixels on the boundary are not considered when connecting,and superpixel segmentation results and grayscale information are applied for three adjacencies.Finally,we do a dot product between the significance image by manifold ranking based on foreground features and the single-pixel logarithmic difference image,and the final binary change image is obtained by threshold segmentation.In this paper,three datasets of dual-phase images are tested.The results indicate that compared with other change detection algorithms,the proposed method can improve the accuracy of change detection effectively.