Fusion-Oriented Aircraft Detection in Large Scene Image Based on Tiny Darknet

Abstract

High resolution and large field of view in large scene images ( $> 10000\times 10000\text{pixels}$ ) provide rich detail information and wide viewing angle. However, the computational complexity is also drastically increased. Small object detection in large scene remote sensing images is always one of the nontrivial problems. Generally, the body length of aircraft in large scene images is only 30∼80 pixels at 0.5m or 1m resolution. While convolutional neural networks (CNN) have made great breakthroughs in natural images, the fast and accurate identification of small objects has not been effectively tackled. In order to improve accuracy, this paper proposes a two-stage Fusion-oriented aircraft detection model (FOADM) and uses an adjusted detection network model Tiny Darknet based on Darknet53 to adapt for the detection of small aircraft. In the first stage, we combine Line Segment Detection (LSD) with Graph-Based Visual Saliency (GBVS) to acquire the region of interest (ROI) in downsampled images, which can reduce the worthless search to some extent. Then, in original high resolution image, the ROI is located and partitioned. In the second stage, each block is fed into the detection network Tiny Darknet for aircraft detection. Experiments demonstrate the proposed method can effectively improve the recall of detection result in high resolution.