Abstract: Abstract: Traditional deep learning-based Remote sensing image object detection methods faced the challenges of high missed rate and inaccurate classification of dense objects. In order to solve these above problems, an anchor-free deep-learning-based object detection method for cluster objects with rotation in remote sensing image was established. Firstly, CenterNet was used as the baseline network, features were extracted from the backbone network, and an angle regression branch was added on the basis of the original network to perform angle regression. Then, a feature enhancement module based on asymmetric convolution was proposed. The feature map extracted from the backbone network was put into the feature enhancement module to reduce the influence caused by the rotation and turnover of the object, then got the center point and size information of the object. The accuracy of this method on the public dataset is 7.80 percentage points higher than Rotation Region Proposal Networks, and 7.50 percentage points higher than the original CenterNet. On the self-built dataset Ship3, the accuracy is 8.68 percentage points higher than Rotation Region Proposal Networks. And the accuracy of this method is 8.47 percentage points higher than the original CenterNet. Compared with the traditional method, the speed is increased by about 25%, the balance between accuracy and speed is reached.

Key words: Deep Learning, Remote Sensing Image, Object Detection, Asymmetric Convolution, Anchor-free Object Detection

摘要： 摘 要: 目前基于深度学习的遥感图像目标检测方法存在密集目标漏检率高、分类不准确的问题。针对上述问题，建立了一种基于深度学习的无锚点的遥感图像任意角度的密集目标检测方法。首先采用CenterNet作为基线模型，经过主干网络提取特征，改造原有检测器结构，加入角度回归分支进行目标角度回归；然后提出一种基于非对称卷积的特征增强模块，将主干网络提取的特征图输入特征增强模块，增强目标的旋转不变性特征，消除由于目标的旋转、翻转带来的影响，进一步提升目标中心点、尺寸信息的回归精度。该方法在公开比赛数据集上正确率比旋转区域候选网络提升了7.80个百分点，比原始CenterNet提升了7.50个百分点，在自建数据集Ship3上正确率比旋转区域候选网络提升了8.68个百分点，比原始CenterNet提升了8.47个百分点。速度相较于传统方法约提升25%，达到了正确率和速度的平衡。

关键词: 深度学习, 遥感图像, 目标检测, 非对称卷积, 无锚点目标检测