针对无人机航拍图像目标检测中视野变化大、时空信息复杂等问题,文中基于YOLOv5(You Only Look Once Version5)架构,提出基于图像低维特征融合的航拍小目标检测模型.引入CA(Coordinate Attention),改进Mo-bileNetV3的反转残差块,增加图像空间维度信息的同时降低模型参数量.改进YOLOv5特征金字塔网络结构,融合浅层网络中的特征图,增加模型对图像低维有效信息的表达能力,进而提升小目标检测精度.同时为了降低航拍图像中复杂背景带来的干扰,引入无参平均注意力模块,同时关注图像的空间注意力与通道注意力;引入VariFocal Loss,降低负样本在训练过程中的权重占比.在VisDrone数据集上的实验验证文中模型的有效性,该模型在有效提升检测精度的同时明显降低复杂度.
Model for Small Object Detection in Aerial Photography Based on Low Dimensional Image Feature Fusion
To address the challenges of significant changes in the field of view and complex spatiotemporal information in unmanned aerial vehicle aerial image target detection,a model for small object detection in aerial photography based on low dimensional image feature fusion is presented grounded on the YOLOv5(you only look once version 5)architecture.Coordinate attention is introduced to improve the inverted residuals of MobileNetV3,thereby increasing the spatial dimension information of images while reducing parameters of the model.The YOLOv5 feature pyramid network structure is improved to incorporate feature images from shallow networks.The ability of the model to represent low-dimensional effective information of images is enhanced,and consequently the detection accuracy of the proposed model for small objects is improved.To reduce the impact of complex background in the image,the parameter-free average attention module is introduced to focus on both spatial attention and channel attention.VariFocal Loss is adopted to reduce the weight proportion of negative samples in the training process.Experiments on VisDrone dataset demonstrate the effectiveness of the proposed model.The detection accuracy is effectively improved while the model complexity is significantly reduced.
You Only Look Once Version 5(YOLOv5)Small Target DetectionAttention Mecha-nismLoss Function
蔡逢煌、张家翔、黄捷
展开 >
福州大学电气工程与自动化学院 福州 350108
You Only Look Once Version5(YOLOv5) 小目标检测 注意力机制 损失函数
万方数据
维普
