基于改进D2Det尺度自适应目标检测算法研究  

Study on Scale Adaptive Target Detection Algorithm Based on Improved D2Det

作　　者：王玲 黄冠 王鹏[1] 白燕娥[1] 邱天衡 WANG Ling;HUANG Guan;WANG Peng;BAI Yane;QIU Tianheng(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学技术学院,长春130022

出　　处：《计算机科学》2023年第S02期174-182,共9页Computer Science

基　　金：中央引导地方科技发展基金吉林省基础研究专项(202002038JC)。

摘　　要：针对D2Det(Towards High Quality Object Detection and Instance Segmentation)面对尺度变化目标和小目标的检测效果不佳并且参数量较大的问题,基于D2Det提出一种尺度自适应的目标检测模型G-SAD2Det。首先在数据预处理阶段引入数据增强算法CutOut和Mosaic,使模型应对复杂场景时有较好的鲁棒性;其次改进特征提取网络ResNet,在每个残差块内构建多尺度特征提取结构,从细粒度层面上更好地提取目标特征,同时在网络结构上添加可切换的全局上下文语义特征提取模块,通过不同池化层来增强显著性特征和全局上下文语义信息;然后改进候选框生成模块,采用自主定位目标中心区域指导候选框的生成,增强算法对尺度变换目标的自适应能力;最后通过Ghost卷积替换普通卷积降低网络的参数量和计算量。使用VOC数据集和COCO子数据集验证算法的有效性,G-SAD2Det比D2Det在两个数据集上的mAP@0.5分别提升了3.6%和4.9%;模型参数量减少了27.42%,计算量减少了35.96%,证明改进后的算法在提高了精度的同时也减少了计算量。Aiming at the problem that D2Det(Towards High Quality Object Detection and Instance Segmentation)has poor detection effect and large parameter quantity in the face of scale change targets and small targets,this paper proposes a scale adaptive target detection model G-SAD2Det based on D2Det.Firstly,in the data preprocessing stage,the data enhancement algorithms CutOut and Mosaic are introduced,and the model has good robustness when dealing with complex scenes.Secondly,the feature extraction network ResNet is improved,the multi-scale feature extraction structure is built in each residual block,and the target features are better extracted from the fine-grained level.At the same time,the switchable global context semantic feature extraction module is added to the network structure,and the salience features and global context semantic information are enhanced through different pooling layers.Then,the candidate frame generation module is improved,and the center area of the self-locating target is used to guide the generation of the candidate frame,so that the adaptive ability of the algorithm to the scaling target can be enhanced.Finally,replacing ordinary convolution with Ghost convolution to reduce the amount of network parameters and computation.VOC data set and COCO sub-data set are used to verify the effectiveness of the algorithm,the mAP@0.5 value of G-SAD2Det increases by 3.6%and 4.9%respectively,compared with D2Det in the two data sets.The number of model parameters reduces by 27.42%and the amount of calculation reduces by 35.96%.It is proved that the improved algorithm not only improves the accuracy,but also reduces the amount of computation.

关 键 词：目标检测 尺度自适应 多尺度特征提取 残差块 区域指导候选框 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]