基于卷积特征建模的目标检测方法  

Object detection based on convolutional feature modeling

作　　者：潘秋羽 王伟[1] 王明明[1] 王道顺[2] Pan Qiuyu;Wang Wei;Wang Mingming;Wang Daoshun(College of Computer Science,Xi′an Polytechnic University,Xi′an 710600,China;Dept.of Computer Science&Technology,Tsinghua University,Beijing 100084,China)

机构地区：[1]西安工程大学计算机科学学院,西安710600 [2]清华大学计算机科学与技术系,北京100084

出　　处：《计算机应用研究》2021年第3期928-931,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61601358,61972225,61902164);陕西省教育厅专项科研计划资助项目(15JK1317);国家新闻出版广电总局数字内容防伪与安全取证重点实验室项目。

摘　　要：现有依赖CNN的目标检测算法常采用特征融合的建模方式来丰富特征表达,虽然该方法一定程度上能有效改善多尺度目标检测,但是在针对复杂场景进行检测时却没有显著的提升。这主要受限于三个问题的影响:长路径特征融合造成的特征间相关性损失;仅设计了单方向的融合连接,忽略了反方向的语义信息弥补;忽略了有效感受野(effective receptive field,ERF)在多尺度检测中的重要性。针对这三点分别设计了二次融合结构(double fusion structure,DFS)、多分支融合模块(multi branch fusion module,MBFM)和感受野增强模块(receptive field enhance module,RFEM)。该方法利用DFS缩短特征层级间的相对路径,然后通过MBFM来同时弥补上层和下层的语义信息缺失,并使用RFEM建模特征通道,增大ERF区域。最终模型在PASCAL VOC 2007测试数据集上达到了85.4%的平均精度均值(mean average precision,mAP),与依赖传统建模方式的检测算法相比,提出的方法提高了2.6%。Existing detection algorithms that rely on CNN often adopt the feature fusion to enrich feature expression.Although the methods can effectively improve multi-scale detection,it has no significant improvement in complex scenes due to the three problems.The loss of feature correlation caused by long-path feature fusion;It complements the semantic information only in single direction.The importance of ERF in multi-scale detection is ignored.This paper proposed DFS,MBFM and RFEM to solve these three problems respectively.The proposed method used DFS to shorten the path within feature maps during the second stage fusion,designed MBFM to made up for the loss of semantic information,and designed RFEM to increase the area of ERF.It makes the mAP value reaching 85.4%in the PASCAL VOC 2007 test dataset.Compared with the detection algorithms depending on the traditional modeling method,the algorithm improves the mAP value of 2.6%.

关 键 词：目标检测 特征相关性 多分支融合 有效感受野 卷积神经网络 

分 类 号：TP391.04[自动化与计算机技术—计算机应用技术]