基于注意力与自适应特征融合机制的小目标检测

Small-Target Detection Based on Attention and Adaptive Feature Fusion Mechanism

随着无人机平台的发展,航拍小目标检测成为当下研究热点。为了更有效地解决航拍小目标检测存在的漏检、错检以及重复检测等问题,提出了一种基于注意力与自适应特征融合机制的小目标检测算法ST-YOLOX(small target-YOLOX)。本算法在CSPDarknet中融合了全局注意力模块(GC)以及可变形卷积(DC),增强主干网络对小目标特征的提取能力;采用四尺度自适应空间特征融合金字塔,抑制不同尺度之间的不一致信息,提升小目标特征表达的准确性;优化损失函数以及标签分配策略,提高算法检测精度。实验表明:ST-YOLOX在VisDrone-DET 2019数据集中的平均检测精度(mAP)为21.83%,比YOLOX-s模型提升了3.78%,比PPYOLOE-s模型提升了2.99%,比YOLOv5-s模型提升了6.21%。航拍结果证明,本文算法的小目标检测准确率得到显著提高。

With the development of UAV,drone-captured scenarios detection has become a hotspot of current research.In order to effectively solve the problem of missing,wrong and repeated detection caused by drone-captured scenarios detection,a novel algorithm named ST-YOLOX based on attention and adaptive feature fusion mechanism is proposed in this article.The algorithm combines the Global Context Module(GC)and Deformable Convolution(DC)in CSPDarknet to enhance the ability of backbone networks of extracting the features from small targets.A four-scale adaptive spatial feature fusion pyramid is used to filter the conflicting information between different scales and improve the expressive accuracy of the small target features.The loss function and label allocation strategies are applied to increase the target detection accuracy.Experiments showed that the mean average precision(mAP)of ST-YOLOX in the VisDrone-DET 2019 dataset reached 21.83%,which was 3.78% higher than that of YOLOX-s prototype,2.99% higher than that of PPYOLOE-s,and 6.21% higher than that of YOLOv5-s.Tests on the actual drone-captured scenarios verified that the accuracy of small-scale target detection was significantly improved.