Fast Aircraft Detection Using End-to-End Fully Convolutional Network

Aircraft detection from remote sensing images of complex background is a challenging task. Existing aircraft detection methods usually consist of two separated stages: proposal generation and window classification, which may be suboptimal for the aircraft detection task. To overcome this shortcoming, we propose a unified aircraft detection framework to simultaneously predict aircraft bounding boxes and class probabilities directly from an arbitrary-sized remote sensing image. Specifically, an end-to-end fully convolutional network (FCN) replaces the fully connected layers in traditional convolutional neural network (CNN). This can greatly reduce the model size while obtaining the comparable detection accuracy. To directly detect aircrafts under multiple scales and different aspect ratios, multiple referenced boxes are introduced. The whole framework can be optimized end-to-end by minimizing a multi-task loss. Extensive experiments on a common dataset demonstrate that the proposed method yields much lower false alarm rates at different recall rates than the state-of-the-art methods, and its speed is more than 35 times faster than the compared methods.