摘要
针对yolov5检测无人机航拍图像时对弱光背景目标监测性能不佳这一问题,基于深度学习方法提出了一种改进算法。首先对目标数据集visdrone2019进行数据归一化,提高后续训练效果。而后对yolov5引入添加了Mish激活函数的动态卷积核和使用分布式偏移卷积替换C3块的C3_DSConv模块,并将上述两种卷积结构融合到yolov5网络中;对网络嵌入Bi-Former注意力机制,提升对小目标检测精度。综上所述,最终得到MODB-yolov5模型,实验结果证明该模型的mAP和recall均有提高,检测阴影、黑暗环境中的车辆时精确度明显上升,且FPS较高,这保证了模型仍可用于快速检测或者实时监测。
An improved algorithm is proposed to address the problem that yolov5 detects UAV aerial images with poor monitoring performance for low light background targets.Firstly,data normalization is performed on the target dataset visdrone2019 to improve the detection effect.Then the dynamic convolution kernel with Mish activation function and the C3_DSConv module using distributed offset convolution to replace the C3 block are introduced,and the above two convolution structures are fused into the yolov5 network;the BiFormer attention mechanism is embedded to improve the accuracy of small target detection.In summary,the MODByolov5 model is finally obtained,and the experimental results prove that the model's mAP and recall are both improved,and the accuracy of detecting vehicles in shadows and dark environments is significantly increased,and the FPS is high,which ensures that the model can still be used for rapid detection or real-time monitoring.
作者
信博夫
XIN Bo-fu(Shenyang Aerospace University,Liaoning 110015,Liaoning)
出处
《电脑与电信》
2024年第1期78-83,共6页
Computer & Telecommunication
