多元信息监督的遥感图像有向目标检测  

Multi-information supervision in optical remote sensing images

作　　者：王家宝 程塨[1] 谢星星 姚艳清 韩军伟[1] WANG Jiabao;CHENG Gong;XIE Xingxing;YAO Yanqing;HAN Junwei(School of Automation,Northwestern Polytechnical University,Xi’an 710129,China)

机构地区：[1]西北工业大学自动化学院,西安710129

出　　处：《遥感学报》2023年第12期2726-2735,共10页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家自然科学基金(编号:61772425);陕西省杰出青年科学基金(编号:2021JC-16)。

摘　　要：遥感图像有向目标检测是遥感图像解译中的一项基础任务,在许多领域有着广泛的应用。由于遥感图像目标尺度差异性大、方向任意且紧密排列,传统目标检测所使用的水平框无法准确的定位目标。因此,遥感图像有向目标检测成为目前遥感领域的研究热点。受益于深度学习的发展,遥感图像有向目标检测取得了突破性进展,但是大多数方法仅在检测头部加入角度预测参数,在训练过程中没有充分利用角度信息和语义信息。本文提出了一种多元信息监督的遥感图像有向目标检测方法。首先,在感兴趣区域提取阶段利用角度信息监督网络学习目标方向,从而使网络第一阶段生成更加贴近遥感图像目标的有向候选区域。其次,为了充分利用图像语义信息,本文在网络第二阶段增加语义分支,并使用图像语义标签进行监督学习。本文以Faster R-CNN OBB为基准,在DOTA数据集上验证所提方法的有效性。本文方法相比基准,平均精度(mAP)提升了2.8%,最终的检测精度(mAP)达到74.6%。Oriented object detection is a basic task in the interpretation of high-resolution remote sensing images.Compared with general detectors,oriented detectors can locate instances with oriented bounding boxes,which are consistent with arbitrary-oriented ground truths in remote sensing images.Currently,oriented object detection has greatly progressed with the development of the convolutional neural network.However,this task is still challenging because of the extreme variation in object scales and arbitrary orientations.Most oriented detectors are evolved from horizontal detectors. They first generate horizontal proposals using the Region Proposal Network (RPN). Then,they classify these proposals into different categories and transform them into oriented bounding boxes. Despite their success, thesedetectors exploit only the annotations at the end of the network and do not fully utilize the angle and semantic information.This work proposes an Angle-based Region Proposal Network (ARPN), which learns the angle of objects and generates orientedproposals. The structure of ARPN is the same as that of RPN. However, for each proposal, instead of outputting four parameters forregression, ARPN generates five parameters, which are the center (x, y), shape (w, h), and angle (t). In the training, we first assign anchorswith ground truths by the Intersection of Unions. Then, we directly supervise the ARPN with the shape and angle information of groundtruths. We also propose a semantic branch to output image semantic results for utilizing the advantage of the semantic information. Thesemantic branch consists of two convolutional layers and is parallel with the detection head. We first assign objects to different scale levelsaccording to their areas. Then, we create semantic labels in each scale and use them to supervise the semantic branch. With the semanticinformation supervision, the model will learn translation-variant features and improve accuracy. Moreover, the outputs of the semanticbranch indicate the objectness in each plac

关 键 词：目标检测 有向目标检测 区域建议提取 多元信息 遥感图像 

分 类 号：TP701[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]