改进的Cascade R-CNN算法在目标检测上的应用  

Application of Improved Cascade R-CNN Algorithm in Target Detection

作　　者：张娜[1] 包梓群 罗源 吴彪 涂小妹 ZHANG Na;BAO Zi-qun;LUO Yuan;WU Biao;TU Xiao-mei(School of Computer and Technology(School of Artificial Intelligence),Zhejiang Science Technology University,Hangzhou,Zhejiang 310018,China;School of Science,Zhejiang Science Technology University,Hangzhou,Zhejiang 310018,China;Zhejiang Guangsha Vocational and Technical University of Construction,Dongyang,Zhejiang 322100,China)

机构地区：[1]浙江理工大学计算机科学与技术学院(人工智能学院),浙江杭州310018 [2]浙江理工大学理学院,浙江杭州310018 [3]浙江广厦建设职业技术大学,浙江东阳322100

出　　处：《电子学报》2023年第4期896-906,共11页Acta Electronica Sinica

基　　金：浙江省重点研发计划项目(No.2020C03094);国家级大学生创新创业训练计划项目(No.202010338024);浙江省教育厅一般科研项目(No.Y202147659)。

摘　　要：针对Cascade R-CNN目标检测算法中存在检测精度较低以及目标遮挡问题,本文提出一种改进的Cas-cade R-CNN网络目标检测算法.该算法在主干网络ResNet101中引入可切换空洞卷积模块(Switchable Atrous Convolu-tion,SAC),该模块主要由两个全局上下文模块以及SAC组件构成,采用SAC组件以不同的空洞卷积率对特征进行卷积,并使用Switch函数收集特征来提高特征提取能力.同时,在ResNet101残差网络中引入坐标注意力机制(Coordi-nate Attention,CA),该机制将位置信息嵌入通道注意力中,用于更好地获取方向感知和位置感知信息,进而提高目标检测精度.此外,针对目标遮挡问题,引入Repulsion Loss损失函数.该损失函数主要由吸引项和排斥项组成,吸引项使得预测框和匹配上的目标框尽可能接近,排斥项使得预测框远离错误目标,进而减少非极大值抑制(Non-Maximum Suppression,NMS)的误检,提高目标检测中遮挡问题的检测精度.实验结果表明,在公开的科大讯飞AI挑战赛数据集上,与原算法测试性能相比,改进的Cascade R-CNN网络对该数据集检出率增长了2.39%,改进算法的识别精度有一定的提高.An improved target detection algorithm based on Cascade R-CNN network is proposed to solve the prob-lems of low detection accuracy and target occlusion in the target detection algorithm based on Cascade R-CNN.The algo-rithm introduces a switchable atrous convolution(SAC)module into the backbone ResNet101,which is composed of two global context modules and SAC components.The SAC component is used to convolution the features with different void convolution rates,and the Switch function is used to collect the features to improve the ability of feature extraction.At the same time,coordinate attention(CA)is introduced in ResNet101 residual network,which embeds position information into channel attention,and is used to obtain direction and position information better to improve the accuracy of target detec-tion.In addition,aiming at the problem of target occlusion,this paper introduces the repulsion loss function,which is main-ly composed of the attraction term and the exclusion term.The attraction term makes the prediction box and the target box on the matching as close as possible,and the exclusion term makes the prediction box away from the wrong target,thereby reducing the false detection of non-maximum suppression(NMS)and improving the detection accuracy of the occlusion problem in object detection.The experimental results show that the detection rate of the improved Cascade R-CNN net-work is 2.39%higher than that of the original algorithm on the open IFLYTEK Challenge dataset,the recognition accuracy of the improved algorithm is improved to a certain extent.

关 键 词：Cascade R-CNN 可切换空洞卷积 Repulsion Loss 目标检测 目标遮挡 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]