基于改进Live-Wire算法的无人机遥感影像标注

标签的制作是深度学习应用的关键步骤,为了克服无人机平台的复杂运动、光照条件不足、地物轮廓复杂等导致遥感影像的地物轮廓提取和标注的难点,文中提出一种改进的Live-wire算法并用于无人机遥感影像的典型地物的标签标注;通过改进模糊隶属度函数克服了Pal-King隶属函数灰度覆盖空间不足的缺陷并结合双阈值方法实现边缘点的提取,以改进的Pal-King的模糊边缘检测方法替代Live-Wire算法的拉普拉斯边缘提取方法;通过增加节点之间梯度幅值的变化特征优化代价函数,以提高Live-Wire算法的轮廓跟踪的连续性;大量的对比实验证明,相较于传统方法,改进的Live-Wire方法的轮廓提取和跟踪的稳健性、效率更高.

Annotation Method of UAV Remote Sensing Images based on Proposed Live-Wire Algorithm

Sample annotation is the key step in the applications of deep learning. In order to solve labeling problems of typical ground objects in UAV remote sensing images caused by the complex motion of UAV platform, insufficient illumination and complexity of object contours. In this paper,an improved live-wire algorithm was proposed and applied into image annotation of typical objects in UAV remote sensing images. The traditional Pal-King fuzzy edge detection method was improved by proposing a fuzzy membership function to over-come defects in gray coverage of traditional PAL-King membership function and the double threshold method was used to extract edge points. Moreover, the proposed fuzzy edge detection method was used to replace the Laplace edge detection algorithm in traditional live-wire method. The cost function was further optimized by increasing change characteristics of gradient amplitude between nodes to improve perfor-mance and continuity of edge detection and contour tracking. A large number of experiments show that the proposed method can detect and track object contours of UAV images with higher robustness and efficien-cy than that of traditional methods.