Aircraft Rotation Detection in Remote Sensing Image Based on Multi-Feature Fusion and Rotation-Aware Anchor

• Published in: Applied sciences 2022-02, Vol.12 (3), p.1291
• Main Authors: Tang, Feifan, Wang, Wei, Li, Jian, Cao, Jiang, Chen, Deli, Jiang, Xin, Xu, Huifang, Du, Yanling
• Format: Article
• Language: English
• Subjects: Accuracy; Air transportation; Aircraft; Aircraft detection; Aircraft performance; Algorithms; Battlefields; Boxes; Classification; Datasets; Deep learning; Feature maps; harris corner detection; Horizontal cells; multi-feature fusion; Nesting; Neural networks; oriented aircraft detection; Remote sensing; retinanet; rotating anchor; Rotation; Semantics; Sensors
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app12031291

Due to the variations of aircraft types, sizes, orientations, and complexity of remote sensing images, it is still difficult to effectively obtain accurate position and type by aircraft detection, which plays an important role in intelligent air transportation and digital battlefield. Current aircraft detection methods often use horizontal detectors, which produce significant redundancy, nesting, and overlap of detection areas and negatively affect the detection performance. To address these difficulties, a framework based on RetinaNet that combines a multi-feature fusion module and a rotating anchors generation mechanism is proposed. Firstly, the multi-feature fusion module mainly realizes feature fusion in two ways. One is to extract multi-scale features by the feature pyramid, and the other is to obtain corner features for each layer of feature map, thereby enriching the feature expression of aircraft. Then, we add a rotating anchor generation mechanism in the middle of the framework to realize the arbitrary orientation detection of aircraft. In the last, the framework connects two sub-networks, one for classifying anchor boxes and the other for regressing anchor boxes to ground-truth aircraft boxes. Compared with state-of-the-art methods by conducting comprehensive experiments on a publicly available dataset to validate the proposed method performance of aircraft detection. The detection precision (P) of proposed method achieves 97.06% on the public dataset, which demonstrates the effectiveness of the proposed method.