Polycentric Circle Pooling in Deep Convolutional Networks for High-Resolution Remote Sensing Image Recognition
Polycentric Circle Pooling in Deep Convolutional Networks for High-Resolution Remote Sensing Image Recognition
复制标题
用于高分辨率遥感图像识别的深度卷积网络中的多中心圆池化
DOI：
10.1109/jstars.2020.2968564
复制
发表时间：
2020
影响因子：
5.5
通讯作者：
Peng Feifei
中科院分区：
文献类型：
--
作者：
Qi Kunlun;Yang Chao;Hu Chuli;Guan Qingfeng;Tian Wenwen;Shen Shengyu;Peng Feifei
大多数现有的基于深度学习的方法使用从卷积神经网络（CNN）中提取的特征图来分类和检测高分辨率遥感图像（HRSI）。然而，直接应用这些功能的分类和目标检测HRSI是有问题的，因为旋转的变化。在这篇文章中，我们使用多中心循环池（PCP）策略来设计网络，以缓解上述问题。PCP网络（PCP-net）结构可以为不同的输入图像尺寸生成固定长度的表示，并编码旋转不变信息。有了这些优点，PCP-网络应该在总体上改进基于CNN的HRSI分类方法。具体而言，在同心圆池化网络结构的基础上，我们改进了结构，使用多个同心圆中心，以产生更强大的旋转不变信息。使用两个具有挑战性的HRSI场景数据集，我们证明了PCP-net提高了CNN架构在场景分类任务中的准确性。PCP-net可以方便地应用于目标检测，因为输出大小是固定的，而不管图像大小。将更快的区域CNN应用于公开的十类对象检测数据集的实验表明，我们提出的PCP可以在HRSI对象检测任务中实现比感兴趣区域池更高的准确性。
Most existing deep learning-based methods use feature maps extracted from convolutional neural networks (CNNs) for classification and detection of high-resolution remote sensing images (HRSIs). However, directly applying these features to classification and object detection in HRSI is problematic because of rotational variations. In this article, we design networks using the polycentric circle pooling (PCP) strategy to alleviate the abovementioned problem. The PCP network (PCP-net) structure can generate a fixed-length representation for different input image sizes and encode rotation-invariant information. With these advantages, PCP-net should in general improve the CNN-based HRSI classification methods. Specifically, on the basis of the concentric circle pooling network structure, we improve the structure using multiple concentric circle centers to generate more robust rotation-invariant information. Using two challenging HRSI scene datasets, we prove that PCP-net improves the accuracy of CNN architectures for a scene classification tasks. PCP-net can be conveniently applied to object detection because the output size is fixed regardless of image size. Experiments applying the faster region-CNN to a publicly available ten-class object detection dataset demonstrate that our proposed PCP can achieve accuracy higher than that of a region of interest pooling in the HRSI object detection task.