一种轻量化YOLOv4的遥感影像桥梁目标检测算法

A lightweight YOLOv4 algorithm for bridge target detection in remote sensing images

深度学习技术发展迅速,在目标检测中表现出良好的适应性。针对YOLOv4算法在遥感影像桥梁目标检测任务中的检测效率较低和模型轻量化不足问题,使用轻量化的MobileNetv3骨干网络替换原始CSPDarkNet53骨干网络,将传统卷积层替换为深度超参数化卷积层(DO_Conv),提出一种兼具精度和检测效率的轻量化模型。实验表明:比较原始YOLOv4算法,本文算法将模型权重降低55%,检测效率提升70%以上,证明了本文改进之处的有效性;在精度方面,本文算法在与SSD、RetinaNet、YOLOv3和CenterNet等经典目标检测算法比较中仍保持精度优势。与YOLOv4算法相比,本文算法在难度较低的检测任务中精度损失较低,但在检测难度较高的DOTA桥梁数据集中精度损失明显。

Deep learning technology develops rapidly and shows good adaptability in target detection.In view of the low detection efficiency and poor model lightweight of YOLOv4 algorithm in bridge target detection task of remote sensing image,this paper proposed a lightweight model with both accuracy and detection efficiency by using the MobileNetv3 backbone network to replace the original CSPDarkNet53 backbone network,and replacing the traditional convolutional layer with the DO_Conv.Experimental results show that compared with the YOLOv4 algorithm,the proposed algorithm reduces the model weight by 55%and improves the detection efficiency by more than 70%,which proves the effectiveness of the proposed improvement.In terms of accuracy,the proposed algorithm still has an advantage over classical target detection algorithms such as SSD,RetinaNet,YOLOv3 and CenterNet.Compared with YOLOv4 algorithm,the accuracy loss is lower in the detection task with lower difficulty,but the accuracy loss is obvious in the DOTA bridge dataset with higher difficulty.How to improve the detection efficiency of the algorithm while maintaining high detection accuracy will be the focus of future work.