多尺度特征融合的目标检测算法  

Object Detection Algorithm Based on Multi-Scale Feature Fusion

作　　者：何婧媛[1] 谢生龙 田原[1] 田琴琴 HE Jingyuan;XIE Shenglong;TIAN Yuan;TIAN Qinqin(College of Mathematics and Computer Science,Yan’an University,Yan’an 716000,Shaanxi China)

机构地区：[1]延安大学数学与计算机科学学院,陕西延安716000

出　　处：《河南科学》2021年第7期1045-1051,共7页Henan Science

基　　金：陕西省教育厅科学研究专项计划项目(19JK0964);延安大学科研计划项目(YDQ2018-05,YDQ2018-08)。

摘　　要：针对SSD目标检测算法在对小目标进行检测时精度与速度不佳,提出一种基于SSD的多尺度特征融合(Multi-scale Feature Fusion Single Shot Detector,MFF-SSD)目标检测模型.该模型以SSD网络为基础,解决了SSD对小目标检测在精度和速度方面存在的问题,将高层特征与低层特征进行融合,并设计融合模块以提取不同尺度特征.对提出的MFF-SSD模型在人脸检测数据集Wider Face上进行测试验证实验.实验结果表明,MFF-SSD模型对目标的检测精度达到78.9%,检测速度达到57 FPS.总体性能更好,充分说明模型在提高目标检测精度的同时也满足了实时性需求.In view of the poor accuracy and speed of SSD target detection algorithm in detecting small targets,a target detection model based on multi-scale feature fusion single shot detector(MFF-SSD)is proposed.Based on SSD network,the model solves the problems of precision and speed in the detection of small targets by SSD.The high level feature and low level feature are fused and fusion modules are designed to extract different scale features.The proposed MFF-SSD model is tested and verified on Wider Face,a face detection data set.The experimental results show that the MFF-SSD model can detect the target with detection accuracy of 78.9%and detection speed of 57 FPS.The overall performance is better,which fully indicates that the model not only improves the target detection accuracy,but also meets the real-time demand.

关 键 词：SSD 多尺度特征融合 目标检测 跳跃连接 人脸检测 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]