改进RetinaNet的混凝土结构损伤视觉检测算法研究  

Research on Visual Inspection Algorithm for Concrete Construction Damage by RetinaNet

作　　者：罗杨 万黎明 李理[1] 刘知贵[1,2] LUO Yang;WAN Liming;LI Li;LIU Zhigui(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621000;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000)

机构地区：[1]西南科技大学计算机科学与技术学院,绵阳621000 [2]西南科技大学信息工程学院,绵阳621000

出　　处：《计算机与数字工程》2023年第11期2654-2658,2735,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:U21A20157)资助。

摘　　要：为了提高混凝土结构损伤视觉检测的自动化水平,提高检测精度,提出了一种基于RetinaNet方法改进的实时目标检测网络。在网络特征提取部分引入了具有位置信息的通道注意力与多头自注意力,多头自注意力模块位于残差网络的最后一个卷积阶段,代替原有的3×3卷积层,而通道注意力则采用嵌入的方式插入到每个残差块中,使每个卷积阶段都能利用注意力机制进行权重调整。改进RetinaNet检测方法的有效性在自建的混凝土结构多类型损伤数据集上进行了验证实验,所提方法的mAP达到了85.0%,检测速度达到了21.9fps,实验结果表明,论文提出的检测方法能够有效进行实时的混凝土结构损伤检测,并且保证了检测精度。A real-time object detection network based on an improved RetinaNet detection method is proposed in order to im⁃prove the automation level of visual detection of concrete structure damage and increase the detection accuracy.Channel attention with location information and multi-head self-attention are introduced in the feature extraction part of the network.The multi-head self-attention module is located in the last convolution stage of the residual network,replacing the original 3×3 convolution layer,while the channel attention is inserted into each residual block,so that the attention mechanism can be used in each convolution stage to adjust the weights.The effectiveness of the improved RetinaNet detection method is verified on a self-constructed multi-type damage dataset for concrete structure.The mAP of the proposed method achieved 85.0%and the detection speed reaches 21.9 fps.The experimental results show that the proposed detection method can effectively perform real-time concrete structure dam⁃age detection and ensure the detection accuracy.

关 键 词：损伤检测 深度学习 目标检测 注意力机制 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]