基于特征融合的自适应多尺度无锚框目标检测算法  

Adaptive multi-scale anchor-free target detection algorithm based on feature fusion

作　　者：熊磊 王凤随[1,2,3] 钱亚萍 Xiong Lei;Wang Fengsui;Qian Yaping(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Anhui Key Laboratory of Detection Technology and Energy Saving Devices,Wuhu 241000,China;Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,芜湖241000 [2]检测技术与节能装置安徽省重点实验室,芜湖241000 [3]高端装备先进感知与智能控制教育部重点实验室,芜湖241000

出　　处：《电子测量与仪器学报》2022年第11期236-244,共9页Journal of Electronic Measurement and Instrumentation

基　　金：安徽省自然科学基金(2108085MF197,1708085MF154);安徽高校省级自然科学研究重点项目(KJ2019A0162);检测技术与节能装置安徽省重点实验室开放基金(DTESD2020B02);安徽工程大学国家自然科学基金预研项目(Xjky2022040);安徽高校研究生科学研究项目(YJS20210448,YJS20210449)资助。

摘　　要：为了提高CenterNet无锚框目标检测网络的目标检测能力,提出一种基于注意力特征融合和多尺度特征提取网络的改进CenterNet目标检测网络。首先,为了提升网络对多尺度目标的表达能力,设计了自适应多尺度特征提取网络,利用空洞卷积对特征图进行重采样获取多尺度特征信息,并在空间维度上进行融合;其次,为了更好地融合语义和尺度不一致的特征,提出了一种基于通道局部注意力的特征融合模块,自适应地学习浅层特征和深层特征之间的融合权重,保留不同感受域的关键特征信息。最后,通过在VOC 2007测试集上对本文算法进行验证,实验结果表明,最终算法的检测精度达到80.94%,相较于基线算法CenterNet提升了3.82%,有效提升了无锚框目标检测算法的最终性能。In order to improve the target detection ability of CenterNet Ancor-free target detection network, an improved CenterNet target detection network based on attention feature fusion and multi-scale feature extraction network was proposed. Firstly, in order to improve the expression ability of the network for multi-scale targets, an adaptive multi-scale feature extraction network was designed. The feature map is resampled by cavity convolution to obtain multi-scale feature information, and the fusion was carried out on the spatial dimension. Secondly, in order to better integrate semantic and scale inconsistent features, a feature fusion module based on channel local attention was proposed. the fusion weight between shallow features and deep features was adaptively learned, and the key feature information of different perceptual domains was retained. Finally, the algorithm was verified on VOC 2007 test set. The experimental results showed that the detection accuracy of the final algorithm reaches 80.94%, which was 3.82% higher than the baseline algorithm CenterNet, and effectively improves the final performance of the Ancor-free target detection algorithm.

关 键 词：目标检测 无锚框 CenterNet 空洞卷积 特征融合 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]