基于SSD及剪枝神经网络的复杂环境下混凝土裂缝识别  

Concrete crack identification in complex environments based on SSD and pruning neural network

作　　者：王燕华[1,2] 何俊泽 张明洲 戴博闻 徐浩然 Wang Yanhua;He Junze;Zhang Mingzhou;Dai Bowen;Xu Haoran(Key Laboratory of Concrete and Prestressed Concrete Structures of Ministry of Education,Southeast University,Nanjing 211189,China;School of Civil Engineering,Southeast University,Nanjing 211189,China;Chien-Shiung Wu College,Southeast University,Nanjing 211189,China)

机构地区：[1]东南大学混凝土及预应力混凝土结构教育部重点实验室,南京211189 [2]东南大学土木工程学院,南京211189 [3]东南大学吴健雄学院,南京211189

出　　处：《Journal of Southeast University(English Edition)》2023年第4期393-399,共7页东南大学学报（英文版）

基　　金：The National Major Scientific Research Instrument Development Project(No.11827801);the National Science and Technology Project(No.2020YFC1511904)。

摘　　要：为解决裂缝识别算法在复杂环境下性能不佳的问题,提出了一种基于单激发多框检测器(SSD)算法的改进方法.该方法通过调整原始SSD算法中不同分辨率先验框数量的组合,实现对存在噪声的裂缝图像的高精度裂缝识别.在真实场景和实验室中采集足够数量的裂缝图像并进行预处理,利用椒盐算法对裂缝数据集添加噪声模拟复杂环境中的裂缝图像.在识别裂缝数据集时,对改进方法与原始SSD算法进行对比分析.结果表明,原始SSD算法和改进方法识别裂缝的准确性均随噪声水平的增加而降低.在高密度下添加20%等级的椒盐噪声时,原始SSD算法识别裂缝的准确率仅为31.7%,而改进方法的准确率则高达93.0%.因此,改进方法具有较强的抗噪能力,可用于复杂环境下的裂缝识别.To solve the problem of poor crack identification algorithm performance in complex environments,an improved method based on a single-shot multibox detector(SSD)algorithm was proposed.This method realized high-precision crack identification for crack images with added noise by adjusting the combination of the number of different resolution prior bounding boxes in the original SSD algorithm.A sufficient number of crack images were captured and preprocessed in actual scenes and laboratories,and noise was added to the crack dataset using a pretzel algorithm to simulate the crack images in complex environments.The improved method was tested along with the original SSD algorithm to identify the crack dataset,and their test results were compared.The results show that the crack identification accuracy of the original SSD algorithm and improved method decreases with increasing noise levels.When a 20%grade of pretzel noise is added at high density,the accuracy in recognizing cracks is 31.7%and 93.0%for the original SSD algorithm and the improved method,respectively.Therefore,the improved method has excellent antinoise ability and can be used for crack identification in complex environments.

关 键 词：裂缝识别 剪枝神经网络 图像加噪 抗噪性能 病害检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]