基于改进SSD的目标检测算法  

Object detection algorithm based on improved SSD

作　　者：彭林聪 王克瑞 周浩[1] 李海燕[1] 余鹏飞[1] PENG Lincong;WANG Kerui;ZHOU Hao;LI Haiyan;YU Pengfei(College of Information,Yunnan University,Kunming 650000,China)

机构地区：[1]云南大学信息学院,昆明650000

出　　处：《激光杂志》2024年第11期71-76,共6页Laser Journal

基　　金：国家自然科学基金(No.62066046);云南省重大科技专项计划(No.202202AD080004)。

摘　　要：针对SSD(Single Shot Multibox Detector)目标检测算法在浅层特征层缺乏语义信息和高层特征层缺乏细节信息造成的漏检、误检等问题,提出了一种改进的SSD目标检测算法。首先引入改进的全面卷积注意力模块CCBAM(Comprehensive Convolutional Block Attention Module),提高网络对于小目标的敏感程度。之后构建分层特征融合网络HFFNet(Hierarchical Feature Fusion Network),使浅层的细节信息和高层的语义信息进行充分的融合,同时在下采样过程中,使用空洞卷积提取不同尺度的特征信息;在上采样过程中,使用像素重组增加高层特征层分辨率的同时减少信息的丢失,之后与低层特征层融合,加强低层特征层的语义信息。最后,采用残差特征融合模块RFFM(Residual Feature Fusion Module),加强高层特征层局部信息和全局信息的整合,同时丰富特征信息。实验表明,在PASCAL VOC2007测试集上mAP@0.5达到79.6%,比原始SSD算法提高了2.4%,检测速度达到了47.8 FPS。Aiming at the problem of missed and false detections caused by the lack of semantic information in shallow feature layers and the lack of detailed information in high-level feature layers in the SSD(Single Shot Multibox Detector)object detection algorithm,an improved SSD object detection algorithm is proposed.First,the improved Comprehensive Convolutional Block Attention Module(CCBAM)is introduced to increase the sensitivity of the network to small targets.Then,a Hierarchical Feature Fusion Network(HFFNet)is constructed to fully integrate the detailed information from shallow layers and the semantic information from high-level layers.In addition,dilated convolutions are used during downsampling to extract features of different scales.In the upsampling process,pixel shuffling is used to increase the resolution of the high-level feature layers while reducing information loss,and then fused with the low-level feature layers to enhance the semantic information in the low-level feature layers.Finally,the Residual Feature Fusion Module(RFFM)is used to improve the integration of local and global information in the high-level feature layers and to enrich the feature information.The experiment demonstrated a 2.4%improvement over the original SSD algorithm,achieving a mAP@0.5 of 79.6%on the PASCAL VOC2007 test set at a recognition speed of 47.8 FPS.

关 键 词：目标检测 CCBAM 特征融合 空洞卷积 RFFM 

分 类 号：TN256[电子电信—物理电子学]