基于旋转框精细定位的遥感目标检测方法研究

遥感图像中的目标往往呈现出任意方向排列, 而常见的目标检测算法均采用水平框检测, 并不能满足这类场景的应用需求. 因此提出一种旋转框检测网络R2-FRCNN. 该网络利用粗调与细调两阶段实现旋转框检测, 粗调阶段将水平框转换为旋转框, 细调阶段进一步优化旋转框的定位. 针对遥感图像存在较多小目标的特点, 提出像素重组金字塔结构, 融合深浅层特征, 提升复杂背景下小目标的检测精度. 此外, 为了在金字塔各层中提取更加有效的特征信息, 在粗调阶段设计一种积分与面积插值法相结合的感兴趣区域特征提取方法, 同时在细调阶段设计旋转框区域特征提取方法. 最后在粗调和细调阶段均采用全连接层与卷积层相结合的预测分支, 并且利用SmoothLn作为网络的回归损失函数, 进一步提升算法性能. 提出的网络在大型遥感数据集DOTA上进行评估, 评估指标平均准确率达到0.7602. 对比实验表明了R2-FRCNN网络的有效性.

Research on Detection Method of Refined Rotated Boxes in Remote Sensing

The objects in remote sensing images are often shown in any direction. The common algorithms of object detection adopt horizontal detection, which cannot fulfill the application requirements in remote sensing. Therefore, this paper proposes an object detector of rotated boxes named R2-FRCNN. The network adopts two stages of rough and refined adjustments to realize the detection of rotated boxes. The rough adjustment stage is used to transform the horizontal boxes into rotated boxes, and the refined adjustment stage is used to further optimize the position of the rotated boxes. In view of the fact that there are many small objects in remote sensing images, this paper proposes a pixel-recombination pyramid structure to improve the detection accuracy of small objects in a complex background by integrating deep and shallow features. In addition, in order to extract more effective feature information from each layer of the pyramid, this paper designs a region pooling method combining integration and area interpolation in the rough adjustment stage, and a region pooling method of rotated boxes in the refined adjustment stage. Finally, this paper adopts the prediction branch combining the fully connected layers and the convolutional layers, and takes the SmoothL_n as the regression loss function of the network to further improve the performance of the algorithm. The network proposed in this paper is evaluated on a large remote sensing dataset DOTA, and the evaluation mean average precision reaches 0.7602. Comparative experiments show the effectiveness of R2-FRCNN modules.