A Path Aggregation Network Based on Residual Feature Enhancement for Object Detection in Remote Sensing Imagery

The complex background and small scale of objects in remote sensing images (RSIs) lead to poor detection performance. To solve these problems, a new object detection method named the residual feature enhancement detection network (RFEDN) is proposed. First, to weaken the interference of irrelevant backgrounds, we introduce a common-and-differential attention network into the cross-stage partial darknet-53, which can refine features in both channel and spatial dimensions so that the network pays attention to meaningful features. After this, we design the residual feature enhancement-based path aggregation network (RFE-PANet), which can help alleviate the problem of information loss caused by channel reduction. We perform experiments on the dataset of object detection in aerial images (DOTA) series datasets and the HRSC2016 dataset. The experimental results reveal the superior performance of the proposed method.