基于双角度多尺度特征融合的无锚框目标检测算法  

Anchor-Free Object Detection Algorithm Based on Dual-Angle Multi-Scale Feature Fusion

作　　者：王小玉[1] 魏钰鑫 芦荐宇 俞越 WANG Xiaoyu;WEI Yuxin;LU Jianyu;YU Yue(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China;School of Measurement-Control Technology and Communication Engineering,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,哈尔滨150080 [2]哈尔滨理工大学测控技术与通信工程学院,哈尔滨150080

出　　处：《哈尔滨理工大学学报》2024年第4期1-9,共9页Journal of Harbin University of Science and Technology

基　　金：国家自然科学基金(61772160);黑龙江省教育厅科学技术研究项目(12541177).

摘　　要：针对无锚框目标检测算法CenterNet中特征利用不充分且检测精度不足的问题,提出一种基于双角度多尺度特征融合的改进算法。首先,通过使用Res2Net网络替换主干网络,使网络从更细粒度的角度提高网络的多尺度表达能力。其次,使用重复加权双向特征金字塔网络从层级角度提升多尺度加权特征的融合能力。最后,加入坐标注意力机制,在避免增加计算资源消耗的前提下增强感受野,将坐标信息嵌入通道注意力中以提升模型对目标的定位提高模型的检测精度。实验结果表明:改进算法在PASCAL VOC数据集和KITTI数据集检测准确率分别达到了82.3%和87.8%,与原CenterNet算法相比精度分别提升5.5%和2.4%。In response to the problem of insufficient feature extraction and insufficient detection accuracy in the anchor-free frame object detection algorithm CenterNet.We propose an improved algorithm based on dual-angle multi-scale feature fusion.Firstly,using a repeated weighted bidirectional feature pyramid network to enhance the fusion ability of multi-scale weighted features from a hierarchical perspective.Secondly,by replacing the backbone network with Res2Net network,the network can improve its multi-scale expression ability from a more fine-grained perspective.Finally,the coordinate attention mechanism is added to enhance the receptive field without consuming a lot of computing resources,and the coordinate information is embedded in the channel attention to improve the model′s target positioning and improve the detection accuracy of the model.The improved algorithm′s detection accuracy in the PASCAL VOC data set and KITTI data set reached 82.3%and 87.8%respectively.Compared with the original CenterNet algorithm,the accuracy increased by 5.5%and 2.4%respectively.

关 键 词：目标检测 注意力机制 无锚框 多尺度特征融合 CenterNet 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]