摘要
混凝土裂缝是影响混凝土结构服役性能的重要因素之一,研究开发一种引入生成对抗网络和注意力机制的SSD目标检测修正算法,对精准识别裂缝目标具有重要意义。通过建立全卷积神经网络与生成对抗网络(GAN)相结合的改进网络结构,以工程采集的有限裂缝样本图像为基础生成高质量图像数据以解决数据集获取难的问题;同时引入混合注意力机制(CBAM)以增强SSD特征提取网络对较小裂缝的感知能力。研究结果表明:使用引入注意力和生成对抗网络的目标检测算法检测混凝土表面裂缝,在裂缝识别时各项性能指标均超过了83%,平均准确率AP更是达到了91.51%,相较于原始的SSD目标检测算法提高了10.36%。
Since concrete cracks are one of the important factors that affect the service performance of concrete structures,it is therefore important to develop an SSD target detection correction algorithm by introducing generative adversarial network and attention mechanisms to achieve accurate identification of crack targets in small samples.An improved network structure is established by combining the full convolutional neural network with the adversarial generative adversarial network(GAN).Based on the improved network structure and limited cracking samples collected from practical projects,high-quality synthetic images are generated.The results showed that the perceptual ability of crack targets can be improved after the mixed attention mechanisms(CBAM)are introduced to modify the SSD target detection.The research results indicate that the average accuracy of the improved algorithm is as high as 91.52%and increases by 10.36%compared with the original.
作者
倪彤元
南晓鹤
刘强
王烨晟
谢政
NI Tongyuan;NAN Xiaohe;LIU Qiang;WANG Yesheng;XIE Zheng(College of Civil Engineering,Zhejiang University of Technology,Hangzhou 310023,China;Zhejiang Key Laboratory ofCivil Engineering Structures&Disaster Prevention and Mitigation Technology,Hangzhou 310023,China;Zhejiang Huadong Mapping and Engineering Safety Technology Co.,Ltd.,Hangzhou 310014,China)
出处
《浙江工业大学学报》
北大核心
2025年第3期237-243,共7页
Journal of Zhejiang University of Technology
基金
国家自然科学基金面上项目(52379136)
浙江省建设科研项目(2023K230)。
关键词
裂缝识别
深度学习
目标检测
生成对抗网络
注意力机制
crack detection
deep learning
object detection
generative adversarial network
attention mechanisms
