基于DATE-FCOS的空中目标检测研究

Research on Air Target Detection based on DATE-FCOS

摘要:

航空图像中目标的检测是当前研究的热点之一,高效、准确的探测在军事和民用领域具有很高的价值。但由于高空环境复杂和空中目标尺度形状多变,不同用途的飞行器进行改装和涂装使得空中目标在检测中难度较大；所以提出了一种改进的一阶端到端的空中目标检测算法；算法采用 DATE-FCOS(Dual Assignment for End-to-End Fully Convolutional One-Stage Object Detection)为基本框架,用CIoU替代GIoU加入到边界框回归损失函数中,并在此基础上,利用可形变卷积模块对其骨干网络进行了改进并且在FPN结构之后加入CBAM(Convolutional Block Attention Module)模块。通过实际实验测试,所提方法在 FGVC aircraft数据集上提高了检测的平均检测精度,达到 77.8%,对比原模型提升11%,满足空中目标检测的应用。

Abstract:

The detection of targets in aerial images is one of the hot spots in current research, and efficient and accurate detection has high value in military and civilian fields. However, due to the complex high-altitude environment and the variable scale and shape of air targets, the modification and coating of aircraft for different purposes make it difficult to detect air targets. Therefore, an improved first-order end-to-end air target detection algorithm is proposed. The algorithm adopts DATE-FCOS (Dual Assignment for End-to-End Fully Convolutional One-Stage Object Detection) as the basic framework, replaces GIoU with CIoU and adds it to the bounding box regression loss function, and on this basis, the deformable convolutional module is used to improve its backbone network and add CBAM ( Convolutional Block Attention Module). Through practical experimental tests, the proposed method improves the average detection accuracy of the FGVC aircraft dataset by 77.8%, which is 11% higher than the original model, which meets the application of aerial target detection.