改进多层尺度特征融合的目标检测算法  

Improved multi-scale feature fusion target detection algorithm

作　　者：李康康 于振中[1,2] 范晓东 宋思远 LI Kang-kang;YU Zhen-zhong;FAN Xiao-dong;SONG Si-yuan(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China;Institute of Intelligent Equipment,HRG International Institute for Research and Innovation,Hefei 230601,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]哈工大机器人国际创新研究院智能装备研究所,安徽合肥230601

出　　处：《计算机工程与设计》2022年第1期157-164,共8页Computer Engineering and Design

基　　金：江苏省自然科学基金项目(BK20130159)。

摘　　要：为提高小目标检测任务的准确率和稳定性,解决SSD(single shot MultiBox detector)算法在小目标识别和定位过程中准确率较低的问题,基于SSD算法提出一种改进方法。在原始的SSD卷积网络结构上进行修改和优化,通过特征图之间的特征融合,重构卷积预测特征图上的物体特征信息。考虑到网络复杂度增加带来的数据分布变化的影响,加入批量归一化BN(BatchNorm)层。在PASCAL VOC2007数据集和生活用品(Supplies Dataset)数据集上的实验结果表明,改进算法的mAP相比原始SSD分别提高了10.4%、15.1%。鉴于网络融合带来的参数增加,改进算法的检测速度仍表现良好,符合算法的实时性要求。To increase the accuracy and stability of target detection task and solve the problem of low accuracy in the process of small target detection and location,an improved method was proposed based on SSD(single shot MultiBox detector)algorithm.The original convolutional network structure was optimized and improved,the object features on different predictive feature maps were increased through feature fusion between feature maps.Considering the influence of the change of data distribution brought by the increase of network complexity,the batch normalized BN(BatchNorm)layer was added.Experimental results on PASCAL VOC2007 dataset and Supplies Dataset show that the mAP of the improved algorithm is improved by 10.4%,15.1%respectively,compared with the original SSD.In view of the parameter increase brought by network convergence,the detection speed of the improved algorithm is not affected too much,which meets the real-time requirements of the algorithm.

关 键 词：深度学习 目标识别与定位 卷积网络 SSD算法 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]