Aircraft Detection in SAR Images via Point Features
Jun Chen, Han Wang, Hao Lu
Abstract
Deep learning-based detection methods have become mainstream to solve aircraft detection problems in SAR images. However, the distinct scattering characteristics displayed by SAR aircraft have been rarely explored in the context of traditional deep learning-based methods. Furthermore, almost all state-of-the-art object detectors, such as YOLO-series, Faster R-CNN, RetinaNet, rely on pre-defined anchor boxes, which may lead to complex computations and redundancy in regard to hyper-parameters. To address these issues, this letter first proposes representing aircraft in SAR images using point features that highlight key semantic areas of the target and are consistent with the scattering properties of SAR imagery. Object areas can then be bounded using the arrangement of these points, avoiding the need to explicitly calculate every anchor box. Further, an Incentive Attention Feature Fusion (IAFF) strategy is proposed to improve feature fusion efficiency and suppress the impact of redundant information on point features. Experiments conducted on four publicly available SAR aircraft detection datasets demonstrate the efficiency and effectiveness of the proposed method.