基于改进的SSD监理目标检测研究  

Research on Supervision Object Detection Based on Improved SSD

作　　者：黄静 谢宣 HUANG Jing;XIE Xuan(School of Information Science and Technology,Zhejiang University of Science and Technology,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息学院,浙江杭州310018

出　　处：《电子科技》2022年第5期7-13,共7页Electronic Science and Technology

基　　金：浙江省重点研发计划(2021C01048)。

摘　　要：针对装饰装修工程中由人工验收带来的诸多问题,文中提出了一种改进的SSD算法并将其应用于监理工作来代替人工验收,推动智能监理的实现。由于SSD算法存在对同一目标复检以及小目标检测效果欠佳等问题,故文中利用DPN网络替换基础特征提取网络VGG16。DPN结合了Resnet和Densenet的优点,具有更好的特征提取能力。通过加权FPN融合特征图,突出不同层特征图的贡献,丰富用于预测的特征图语义。利用深度可分离卷积降低模型的参数量,提高算法的推理速度。实验对比发现,改进后模型的平均精度提升了3.47%,对小数目检测平均精度的提升可达15%,证明新模型在监理目标检测任务中效果良好。In view of many problems caused by manual acceptance in decoration projects,this study proposes an improved SSD algorithm and applies it to supervision work to replace manual acceptance and promote the realization of intelligent supervision.Because the SSD algorithm has problems such as rechecking the same target and poor detection of small targets,the DPN network is employed to replace the basic feature extraction network VGG16.DPN combines the advantages of Resnet and Densenet,and has better feature extraction capabilities.Feature maps are fused by weighted FPN to highlight the contributions of feature maps of different layers and enrich the semantics of feature maps for prediction.Using depth separable convolution can reduce the amount of model parameters and improve the inference speed of the algorithm.Experimental comparison shows that the average accuracy of the improved model is increased by 3.47%,and the average accuracy of small numbers of detection is increased by up to 15%,which proves that the new model is effective in the task of supervision target detection.

关 键 词：监理 SSD VGG16 DPN Resnet Densenet 深度可分离卷积 FPN 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论] TN99[自动化与计算机技术—计算机科学与技术]