基于尺度感知的多路径特征融合目标检测  

Multi-path Feature Fusion Object Detection Based on Scale-aware

作　　者：潘浩 郑华 陈清俊[1] 廖晓琦[1] 王泓楷 PAN Hao;ZHENG Hua;CHEN Qing-Jun;LIAO Xiao-Qi;WANG Hong-Kai(College of Photonic and Electronic Engineering,Fujian Normal University,Fuzhou 350007,China;Key Laboratory of Optoelectronic Science and Technology for Medicine(Ministry of Education),Fujian Normal University,Fuzhou 350007,China;Fujian Provincial Key Laboratory of Photonic Technology,Fujian Normal University,Fuzhou 350007,China;Fujian Provincial Engineering Technology Research Center of Photoelectric Sensing Application,Fujian Normal University,Fuzhou 350007,China)

机构地区：[1]福建师范大学光电与信息工程学院,福州350007 [2]福建师范大学医学光电科学与技术教育部重点实验室,福州350007 [3]福建师范大学福建省光子技术重点实验室,福州350007 [4]福建师范大学福建省光电传感应用工程技术研究中心,福州350007

出　　处：《计算机系统应用》2022年第12期251-258,共8页Computer Systems & Applications

基　　金：福建省高校产学合作项目(2021H6025)

摘　　要：在通用的目标检测算法中,目标多变的尺度和特征融合利用一直是限制目标检测任务的难题.针对上述问题,首先文中提出了多路径特征融合模块,模块采用跨尺度跨路径特征融合的方法,强化输入输出特征之间的联系,缓解了特征信息在传递时的稀释问题.同时,文中通过改进注意力模型提出了尺度感知模块,该模块能根据目标的尺度自行地选择感受野大小,从而使模型易于识别多尺度目标.将尺度感知模块嵌入到多路径特征融合模块中,使模型的特征提取和利用能力均得到提升.经实验验证,文中提出的算法在数据集PASCAL VOC和MS COCO上的平均检测精度分别达到了82.2%和38.0%,相比基线FPN Faster RCNN分别提升了1.3%和0.6%,其中对小尺度目标的检测效果提升最为显著.The variable scales of objects and the use of feature fusion have been the challenges for popular object detection algorithms.Considering the problems,this study proposes a multi-path feature fusion module,which strengthens the connection between input and output features and alleviates the dilution of feature information in transmission by adopting cross-scale and cross-path feature fusion.Meanwhile,the study also proposes a scale-aware module by refining the attention model,which allows the model to easily recognize multi-scale objects by selecting the size of the receptive field corresponding to the scale of the objects independently.After the scale-aware module is embedded into the multipath feature fusion module,the feature extraction and utilization abilities of the model are improved.The experimental results reveal that the proposed method achieves 82.2 mAP and 38.0 AP on PASCAL VOC and MS COCO datasets,respectively,an improvement of 1.3 mAP and 0.6 AP over the baseline FPN Faster RCNN,respectively,with the most significant improvement in detection of small-scale objects.

关 键 词：目标检测 特征融合 注意力机制 尺度感知 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]