Small-scale aircraft detection in remote sensing images based on Faster-RCNN

Detecting aircraft in remote sensing images becomes increasingly important in both military and civilian fields. However, the accuracy of existing detection approach is not high enough especially for the small-scale aircraft when considering the size and scenario of the remote sensing images. To improve the accuracy of detecting small-scale aircraft, this paper proposes a detection approach for aircraft based on Faster-RCNN, called MFRC. Firstly, the K-means algorithm is used to cluster aircraft data in remote sensing images. Anchors are improved based on clustering results. Secondly, to extract location features of small-scale aircraft, the layer of pooling in the VGG16 network is reduced from four to two. Finally, the Soft-NMS algorithm is used to optimize the aircraft bounding boxes. In the experimentation, MFRC is evaluated under different conditions and compared with other models. The experimental results show that MFRC can detect small-scale aircraft effectively and the accuracy is improved by 3% compared to existing methods.