摘要
为提升原始SSD算法的小目标检测精度及鲁棒性,提出一种基于通道注意力机制的SSD目标检测算法。在原始SSD算法的基础上对高层特征图进行全局池化操作,结合通道注意力机制增强高层特征图的语义信息,并利用膨胀卷积结构对低层特征图进行下采样扩大其感受野以增加细节与位置信息,再通过级联的方式将低层特征图与高层特征图相融合,从而实现小目标及遮挡目标的有效识别。实验结果表明,与原始SSD算法相比,该算法在PASCAL VOC数据集上的平均精度均值提升了2.2%,具有更高的小目标检测精度和更好的鲁棒性。
In order to improve the accuracy and robustness of the original Single Shot Multibox Detector(SSD)algorithm for small target detection,this paper proposes a SSD target detection algorithm based on channel attention mechanism.The algorithm implements global pooling on high-level feature maps on the basis of the original SSD algorithm.Then the semantic information of high-level feature maps is enhanced by using the channel attention mechanism,and the dilated convolution structure is introduced for subsampling of low-level feature maps to enlarge their receptive field,so as to improve details and location information.Finally,different levels of feature maps are fused by cascading to implement effective recognition of small objects and occluded objects.Experimental results on the PASCAL VOC dataset show that compared with the original SSD algorithm,the proposed algorithm improves the mean Average Precision(mAP)by 2.2%with a higher accuracy and robustness for small object detection.
作者
张海涛
张梦
ZHANG Haitao;ZHANG Meng(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2020年第8期264-270,共7页
Computer Engineering
基金
辽宁省自然科学基金(20170540426)。
