Abstract: In forestry intelligence applications, it is necessary to identify non-rigid objects, such as fire, smoke or cloud in remote sensing image. Limit to object particularity in shapelessness or color-uncertainty, these lead to poor performance or even failure of existing image annotation methods. Therefore, how to accurately select the region of interest（ROI） of the target object is studied to improve the speed and accuracy of pixel annotation. Aiming at the non-convexity of the image object shape, a pixel-level automatic labeling algorithm for ROI of any shape is proposed. Pixels are extracted, trained, and labeled. Based on high-definition images captured by a drone, comparing with the bounding box, the convex hull, image segmentation and image matting methods, the results show the superiorities in easy operation, rapidity and pixel-level precision, conforms to human vision, and can achieve automatic pixel-level annotation of non-rigid objects.

Key words: image annotation, regions of interest, convex hull, arbitrary-shaped

摘要： 在林业智能应用问题中，经常需要对遥感图像中的非刚性目标如“火”“烟”“云”等进行识别。由于此类目标物受限于颜色不确定（或渐变）以及无固定形状，导致现有图像标注方法的性能不佳甚至失效。因此，针对如何准确地选择目标物的兴趣区域（region of interest，ROI），进而提高像素标记的速度以及准确率进行了研究。针对图像目标物形状的非凸性，提出了任意形状ROI的像素级自动标注算法，该算法可将任意形状的兴趣区域转化为多个凸区域问题，针对分解后的每个凸壳内的像素进行提取、训练、标记。基于无人机拍摄的高清图像集，通过将其与边界框法、凸壳法、图像分割以及图像抠图方法进行实验对比，证明该方法存在易操作，符合人类视觉，求解速度快，可批量操作，且能够实现非刚性目标的像素级自动标注等特点。

关键词: 图像标注, 兴趣区域, 凸壳, 任意形状