Abstract:

Aiming at the shortcomings of the general target detection framework for insufficient target detection performance in aerial imagery, an improved Mask R-CNN algorithm is proposed for target detection in aerial imagery. The algorithm adds an image fusion network to fuse visible light and infrared images to eliminate the impact of shadowing on the target. At the same time, the feature pyramid structure is improved, so that the high-level semantic features and low-level localization information in the feature extraction process are fully integrated, and the detection accuracy of targets at various scales is improved. In order to solve the problems of low detection accuracy and high localization difficulty of small targets, the algorithm uses a new region proposal network SD-RPN, and sets a reasonable size sliding window on convolutional layers of different depths to detect targets of different scale types, which makes the proposed region more accurate. Experimental results show that the proposed algorithm performs well on the VEDAI dataset in comparison with current mainstream detection algorithms, and the detection accuracy is greatly improved, especially for small target detection.

Key words: aerial image, target detection, image fusion, feature pyramid, region proposal network

摘要：

针对通用目标检测算法在检测航空影像目标所表现的性能缺陷，提出一种改进Mask R-CNN算法用于航空影像的目标检测。该算法增加图像融合网络，将可见光图像与红外图像进行融合，消除目标被阴影遮蔽对检测造成的影响；同时改进了特征金字塔结构，使特征提取过程中的高层语义特征和低层定位信息得到充分融合，各尺度目标的检测精度得到提升；为解决小目标检测精度低和定位难度高的问题，该算法采用新型区域建议网络SD-RPN，在不同深度的卷积层设置合理大小的滑动窗口，用以检测不同尺度类型目标，使建议区域更加精准。实验结果表明，相比较主流检测算法，该算法在VEDAI数据集上表现出色，检测精度提升较大，尤其是小目标检测的精度提升显著。

关键词: 航空影像, 目标检测, 图像融合, 特征金字塔, 区域建议网络