多尺度特征融合的目标检测算法研究  

Research on target detection algorithm based on multi-scale feature fusion

作　　者：王君[1] 陈敏 董明利[1] 燕必希[1] WANG Jun;CHEN Min;DONG Mingli;YAN Bixi(School of Instrument Science and Opto Electronic Engineering,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学仪器科学与光电工程学院,北京100192

出　　处：《光学技术》2022年第6期749-754,共6页Optical Technique

基　　金：教育部“长江学者与创新团队发展计划”(IRT_16R07);北京信息科技大学2021年人才培养质量提高项目(5102110802)。

摘　　要：针对Faster R-CNN在多尺度目标检测时易出现小目标漏检和误检的问题,提出一种改进的多尺度目标检测算法。将利于小目标检测的低层网络和利于大尺度目标检测的高层网络进行多尺度特征融合;在训练阶段,采用在线难例样本挖掘算法维护难例样本分类池,加速神经网络模型迭代收敛,解决训练样本不均衡、训练效率低下的问题;计算并统计待检测目标的尺度大小,合理控制用于生成候选区域的锚框尺寸,提高模型泛化能力。采用PASCAL VOC2012公开数据集和类人足球机器人自建数据集进行算法验证,实验结果表明,相比Faster R-CNN算法,本算法的平均检测精度在上述数据集下分别提高了8.61和5.47个百分点。Aiming at the problem that Faster R-CNN is prone to miss detection of small targets in multi-scale target detection,an improved multi-scale target detection algorithm is proposed.Firstly,the low-level features used for small target detection are fused with high-level features used for large target detection.Secondly,an online hard example mining algorithm is applied to maintain a hard case sample classification pool for accelerating the iterative convergence of the neural network model,which could solve the problems includes uneven training samples and low training efficiency.Finally,the size of the detected target is calculated and counted for controlling the size of the anchor box and improving the generalization ability of the model.The experimental results show that compared with Faster R-CNN,the mean average precision(mAP) is improved by 8.61 and 5.47 percentage points respectively.

关 键 词：目标检测 多尺度特征融合 在线样本难例挖掘 Faster R-CNN 足球机器人 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP242.6[自动化与计算机技术—计算机科学与技术]