Abstract: Faster region-based convolutional neural network (Faster RCNN), a two-stage object detection model, proposes candidate regions through a region proposal network and coalesces the two stages of region proposal and classification to a single network, which makes most computation be conducted in the graphic processing unit. Therefore, it has both high detection speed and accuracy on many public datasets. However, when we directly use Faster RCNN to conduct target detection on the remote sensing imageries, its performance is not ideal enough. In this work, the influences of pooling and target scale on region proposal are analyzed, and an optimized region proposal network is proposed to address the problems in the region proposal process. In addition, an individualized generation strategy for training samples which can avoid the generation of invalid foreground samples is introduced to speed up the training process. We evaluate the proposed detection method on the remote sensing image dataset. The result shows that the proposed model improves the recall rate and detection accuracy of multi-scale targets. Moreover, the experiments also demonstrate that the training of the model is rapid and high-efficient.

Key words: remote sensing imagery, target detection, convolutional neural network (CNN), deep learning