摘要
针对传统SSD算法在检测小目标时容易漏检且检测精度不高的问题,提出一种改进算法。对SSD算法中各尺度特征进行尺寸大小不变的卷积操作,将卷积前后对应的特征进行轻量级网络融合,从而生成新的金字塔特征层,并加入基于残差模块的检测单元避免增加网络模型容量和运算复杂度,同时增强对小尺度目标的检测能力。基于PASCAL-VOC2007小目标数据集的实验结果表明,与传统SSD、YOLOv3、Faster RCNN等算法相比,在PASCAL-VOC2007小目标数据集中,该算法的m AP指标较传统SSD算法提高8.5%,较Faster RCNN算法提高3.9%,较YOLOv3提高2%,FPS达到83 frame/s,其检测地铁安检图片的m AP达到77.8%。
In the detection of small objects,the traditional SSD algorithm tends to miss the targets,and its detection accuracy is reduced.To address the problem,an improved algorithm is proposed.Each scale feature in the SSD algorithm is convoluted with the size being unchanged,and the corresponding features before and after the convolution are fused by using a lightweight network to generate a new set of pyramid feature layers.Then a detection unit based on the residual module is added to avoid increasing the network model capacity and computational complexity,and to enhance the detection ability of small targets.The experimental results on the PASCAL-VOC2007 small target data set show that,compared with traditional SSD,YOLOv3,Faster RCNN algorithms etc.,the mAP of the proposed algorithm is 8.5%higher than that of the traditional SSD,3.9%higher than that of the Faster RCNN algorithm,and 2%higher than that of the YOLOv3 algorithm.The proposed algorithm increases the FPS to 83 frames/s,and the mAP of the subway security image check to 77.8%.
作者
张震
李孟洲
李浩方
马军强
ZHANG Zhen;LI Mengzhou;LI Haofang;MA Junqiang(School of Electrical Engineering,Zhengzhou University,Zhengzhou 450001,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2021年第7期314-320,共7页
Computer Engineering
基金
国家重点研发计划“公共安全风险防控与应急技术装备”重点专项(2018YFC0824XXX)。
关键词
SSD算法
网络融合
金字塔特征层
残差模块
检测单元
目标检测
SSD algorithm
network convergence
pyramid feature layer
residual module
detection unit
object detection
