基于改进SSD和ISESAM融合的无人机弱小目标检测方法

Detection of UAV Weak and Small Objects Based on Improved SSD and ISESAM Fusion

针对低空无人机目标检测过程中出现的弱小目标、飞行模糊和背景相似,导致目标检测精度降低的问题,论文提出了改进SSD算法以提升目标检测精度。采用轻量化SSD(SSD-MobileNetV3)算法为检测框架,减少模型参数,降低计算复杂度。同时提出一种改进时空注意力(ISESAM)模块,构建通道和空间注意力机制,嵌入YOLOV4网络提高弱小目标检测精度。改进SSD算法的激活函数有效解决神经元可能“失活”的问题,可在低层网络中更有效地提取图像特征。在PASCAL VOC和COCO合成的数据集上实验,结果表明论文提出的改进SSD算法获得的检测精度为90.0%。与SSD-MobileNetV3相比,检测精度提高了4.0%,说明论文提出的改进SSD方法对低空无人机目标的检测更加具有鲁棒性。

In the problems of weak targets,flight blur and background similarity in the process of low-altitude UAV target detection,an improved SSD algorithm is proposed to improve the target detection accuracy.The lightweight SSD(SSD-MobileNetV3)algorithm is used as the detection framework to achieve low-altitude UAV target detection,reducing model parameters and computational complexity.Simultaneously,an improved spatiotemporal attention(ISESAM)module is proposed to construct channel and spatial attention mechanisms to improve the detection accuracy of weak and small targets.The activation function of the improved SSD algorithm can effectively solve the problem of possible"inactivation"of neurons and can more effectively extract image features from the low-level network.Experiments on the datasets synthesized by PASCAL VOC and COCO show that the detection accuracy obtained by the method proposed in this paper is 90.0%.Compared with SSD-MobileNetV3,the detection accuracy is improved by 4.0%,indicating that the improved SSD method proposed in this paper is more robust for the detection of low-altitude UAV targets.