One important area of research in the field of remote sensing information processing is image object detection. Finding expected objects in an image and returning their category confidence and bounding boxes is the objective of the visual object detection task. Remote sensing images, in contrast to conventional images, contain many objects with large aspect ratios and are densely distributed, posing numerous challenging issues. In this paper, we propose Inception-Det, which makes use of a two-stage detection head design to solve these issues. The first stage is used to predict rotational anchors close to the GT-Box, and the second stage uses a higher positive IoU threshold and complete features calculated by FPN Inception to get better detection results. By effectively resolving the issue of objects with large aspect ratios being difficult to regress to high-precision bounding boxes, our proposed feature-complete transform (FCT) can effectively extend the detailed information contained in the feature map without introducing background noise. Extensive testing on two publicly available datasets, DOTA and HRSC2016, demonstrates that our proposed method outperforms the alternatives and further enhances detection performance at a rapid rate.

遥感信息处理领域的一个重要研究领域是图像目标检测。在图像中找到预期的对象并返回它们的类别置信度和边界框是视觉对象检测任务的目标。与传统图像相比，遥感图像包含许多具有大纵横比且分布密集的对象，提出了许多具有挑战性的问题。在本文中，我们提出了 Inception-Det，它利用两级检测头设计来解决这些问题。第一阶段用于预测靠近GT-Box的旋转锚点，第二阶段使用更高的正IoU阈值和FPN Inception计算的完整特征以获得更好的检测结果。通过有效解决宽高比大的对象难以回归到高精度边界框的问题，我们提出的特征完全变换（FCT）可以有效地扩展特征图中包含的详细信息，而不会引入背景噪声。对两个公开可用的数据集 DOTA 和 HRSC2016 的广泛测试表明，我们提出的方法优于替代方法，并进一步快速提高了检测性能。