摘要
针对三维场景存在多样化多尺度目标等问题,提出了一种基于数据融合和双向特征金字塔的多尺度目标检测网络算法。首先,构造了一种基于通道注意力残差结构改进卷积特征的提取网络,并引入注意力机制自适应提取目标融合特征,从而构建了基于数据融合和引入注意力机制残差结构的目标检测网络。接下来,为了使网络更适配多尺度目标,利用双向特征金字塔结构增强融合不同维度特征,并划分不同单元对多尺度目标进行检测,以提高多尺度目标检测的鲁棒性。实验结果表明,与相关算法相比,所提算法对于中小目标的检测准确率较高。
Aiming at the problems of diverse and multi-scale targets in 3D scenes,a target detection network based on data fusion and attention residual structure is proposed.First,a channel attention residual structure is proposed to improve the convolutional feature extraction network,and an attention mechanism is introduced to adaptively extract target fusion features,thereby constructing a target detection network based on data fusion and attention residual structure.Next,in order to make the network more suitable for multi-scale targets,the bidirectional feature pyramid structure is adopted to enhance the fusion of features of different dimensions,and different units are divided to detect the multi-scale targets,which improved the robustness of multi-scale target detection.Experimental results show that the proposed algorithm has a higher detection accuracy for small and medium-sized targets compared to related algorithms.
作者
张浩钧
郭建峰
张锦忠
柴博松
侯诗梦
张学成
ZHANG Haojun;GUO Jianfeng;ZHANG Jinzhong;CHAI Bosong;HOU Shimeng;ZHANG Xuecheng(Shanghai Radio Equipment Research Institute,Shanghai 201109,China;School of Economics and Management,Xi an University of Posts and Telecommunications,Xi an 710061,China;School of Software,Northwestern Polytechnical University,Xi an 710072,China;Third Military Representative Office of Army Equipment Department in Shanghai,Shanghai 200032,China)
出处
《西安邮电大学学报》
2023年第1期92-103,共12页
Journal of Xi’an University of Posts and Telecommunications
基金
上海市军民融合项目(2019-jmrh1-kj18)
陕西省三秦学者岗位“计算金融与风险管理”项目。