ABSTRACT

Detecting geospatial objects, especially small, time-sensitive targets such as airplanes and ships in cluttered scenes, is a substantial challenge in large-scale, high-resolution optical satellite images. Directly detecting targets in countless image blocks results in higher false alarms and is also inefficient. In this paper, we introduce a hierarchical architecture to quickly locate related areas and detect these targets effectively. In the coarse layer, we use an improved saliency detection model that utilizes geospatial priors and multi-level saliency features to probe suspected regions in broad and complicated remote sensing images. Then, in the fine layer of each region, an efficacious end-to-end neural network that predicts the categories and locations of the objects is adopted. To improve the detection performance, an enhanced network, adaptive multi-scale anchors, and an improved loss function are designed to overcome the great diversity and complexity of backgrounds and targets. The experimental results obtained for both a public dataset and our collected images validated the effectiveness of our proposed method. In particular, for large-scale images (more than 500 km²), the adopted method far surpasses most unsupervised saliency models in terms of the performance in region saliency detection and can quickly detect targets within 1 minute, with 95.0% recall and 93.2% precision rates on average.

摘要

在大规模，高分辨率的光学卫星图像中，检测地理空间物体，尤其是时间敏感的小型目标（例如飞机和轮船）是一项巨大的挑战。直接检测无数图像块中的目标会导致更高的误报率，并且效率也很低。在本文中，我们介绍了一种层次结构，可以快速定位相关区域并有效地检测这些目标。在粗糙层中，我们使用改进的显着性检测模型，该模型利用地理空间先验和多层显着性特征来探测宽而复杂的遥感图像中的可疑区域。然后，在每个区域的精细层中，采用有效的端到端神经网络来预测对象的类别和位置。为了提高检测性能，增强了网络，自适应多尺度锚点和改进的损失函数旨在克服背景和目标的巨大多样性和复杂性。从公共数据集和我们收集的图像获得的实验结果验证了我们提出的方法的有效性。特别是对于大型图像（超过500 km²），在区域显着性检测方面，所采用的方法远远超过了大多数无监督显着性模型，可以在1分钟内快速检测目标，召回率平均为95.0％，准确率达到93.2％。