双尺度网络高分辨率楼面影像微小缺陷检测  

Detection of small defects on a building wall surface from high-resolution images using dual-scale neural networks

作　　者：孙光民 陈佳阳 李冰 李煜 闫冬 SUN Guangmin;CHEN Jiayang;LI Bing;LI Yu;YAN Dong(Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China;China National Tobacco Corporation Beijing Corporation, Beijing 100020, China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]中国烟草总公司北京市公司,北京100020

出　　处：《哈尔滨工程大学学报》2021年第2期286-293,共8页Journal of Harbin Engineering University

基　　金：国家重点研发计划(2018YFF01012300);国家自然科学基金项目(11527801,41706201).

摘　　要：为了便于对建筑外墙瓷砖松动和开裂现象进行定期排查以保证周围居民的人身安全,本文提出了一种通过高分辨率相机拍摄的楼面图像进行微小缺陷自动检测的方法。将原始检测任务划分为大尺度下的非墙体分割任务以及小尺度下的缺陷检测任务;分别针对这些任务训练相应的深度模型并应用其进行处理;将这些多尺度任务的处理结果进行融合,得到微小缺陷的最终检测结果。实验表明:本文算法在精度和效率上都要明显优于单尺度方法。本文算法已在某小区实际部署运行并取得了良好的效果,具有很高的实用价值。To facilitate the regular monitoring of exterior building walls for ensuring the personal safety of residents living near the building,we propose a method for automatically detecting small defects via images of the building wall surface captured using high-resolution cameras.With this method,the risks caused by loosening or cracking tiles can be easily identified.First,the original detection task is divided between a large-scale segmentation task of non-tile regions and the small-scale detection of defects.Second,corresponding low-resolution deep models are trained and applied to these tasks.Lastly,the results obtained from these multiscale tasks are fused to obtain the comprehensive detection of small defects.Our experimental results indicate that the accuracy and efficiency of the proposed algorithm are superior to those of the single-scale method.The proposed method has achieved excellent results in real-world applications in a residential area,which confirms its high practical value.

关 键 词：目标检测 墙面 缺陷 高分辨率检测器 卷积神经网络 多尺度 滑窗 负反馈技术 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]