Abstract: Satellite remote sensing images have the characteristics of complex background, different object scales, different directions, and unclear texture. The mainstream object detection algorithms based on deep learning cannot be directly applied to target detection in satellite remote sensing images. RetinaNet was improved to make it suitable for satellite remote sensing images. First, a new feature fusion method was designed to fuse the feature maps output by ResNet50, so that the fused feature maps have both highlevel semantic information and lowlevel texture detail information. Then, in order to reduce the influence of the complex background of the remote sensing image on the object features, a feature perception module was designed to reduce the influence of noise on the feature map while enhancing the useful features. Images of ship, plane and storagetank in the DOTA dataset were selected for training and testing. Experimental results show that, compared with RetinaNet, the improved algorithm has increased the average accuracy of plane, ship and storagetank by 4.1%, 2.5% and 2.4% respectively. Experimental results based on real image data of GF-2 satellite show that the proposed algorithm can be used to intelligently extract multiple types of targets in island and reef images from satellite remote sensing.

Key words: deep learning, remote sensing imagery, object detection, feature fusion, feature perception module, DOTA dataset