基于单阶段目标检测算法的混凝土裂缝识别模型  

A Model for Concrete Crack Recognition Based on One-stage Object Detection Algorithm

作　　者：石子 吴志刚 胡继峰 甘元楠 苏敏 强晟 SHI Zi;WU Zhi-gang;HU Ji-feng;GAN Yuan-nan;SU Min;QIANG Sheng(College of Water Conservancy&Hydropower Engineering,Hohai University,Nanjing 210098,China;China Anneng Group Second Engineering Bureau Co.,Ltd.,Nanchang 330000,China;Zhejiang Water Conservancy and Hydropower Survey and Design Institute Co.,Ltd.,Hangzhou 310002,China)

机构地区：[1]河海大学水利水电学院,江苏南京210098 [2]中国安能集团第二工程局有限公司,江西南昌330000 [3]浙江省水利水电勘测设计院有限责任公司,浙江杭州310002

出　　处：《水电能源科学》2025年第2期118-122,共5页Water Resources and Power

基　　金：国家自然科学基金项目(52079049)。

摘　　要：混凝土结构产生裂缝会严重影响建筑物的安全稳定运行。为实现混凝土出露裂缝的实时高效检测,提出了一种新的单阶段混凝土裂缝检测模型CrackDetectX,该模型由基础特征提取网络、多级特征融合网络、检测头三部分组成。采用动态蛇形卷积(DSConv)精准捕捉裂缝特征,添加反向残差注意力模块(iRMB)融合不同尺度的上下文信息,使神经网络能够为特征图产生更好的像素级注意力。在检测头中引入一种基于MPDIoU的损失函数,全面考虑边界框所有信息,使模型更好地处理边界框宽度及高度的差异。此外,在模型中还引入Lion优化器保存动量信息,并利用其独特的更新规则来更新梯度,提高模型训练的效率。最后对所建模型进行评估,平均精度AP_0.5/%、AP_0.5-0.95/%、参数量、计算量及推理速度分别为93.1%、77.8%、1.62 M、4.3 GFLOPs和61.4 FPS,均优于对比方法,表明所提模型具有良好的鲁棒性,是一种高精度、高效率、轻量化的混凝土出露裂缝检测方法。Cracks in concrete structures can severely affect the safe and stable operation of buildings.To facilitate real-time and efficient detection of visible cracks in concrete,this paper put forward a new single-stage concrete crack detection model,named CrackDetectX,composed of three main parts:a basic feature extraction network,a multi-level feature fusion network,and detection heads.The model employs dynamic snake convolution(DSConv)to precisely capture crack characteristics and integrates an inverted residual mobile block(iRMB)that fuses context information across different scales,providing enhanced pixel-level attention in the feature maps.The detection head incorporates a loss function based on MPDIoU that comprehensively considers all bounding box information,allowing the model to better address variations in their widths and heights.Additionally,the model incorporates the lion optimizer that requires only the retention of momentum information and uses its unique rule for gradient updates,improving the training efficiency of the model.Upon evaluation,the model achieved average precisions of 93.1%for AP_0.5%and 77.8%for AP_0.5-0.95%,with a parameter count of 1.62 M,a computational demand of 4.3 GFLOPs,and an inference speed of 61.4 FPS,all of which surpass the compared methods.The test results show that the proposed model exhibits good robustness and is a high-precision,high-efficiency,and lightweight model for detecting exposed cracks in concrete.

关 键 词：裂缝检测 动态蛇形卷积 反向残差注意力 MPDIoU Lion优化器 CrackDetectX 

分 类 号：TV544[水利工程—水利水电工程] TV698.1