考虑完整性分割的超轻量化路面裂缝检测方法
Ultra-lightweight Pavement Crack Detection Method Considering Complete Segmentation
梁晓 1邵天义 2王雪玮 3李韶华 3郭京波 4申永军1
作者信息
- 1. 石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,河北石家庄 050043;石家庄铁道大学 机械工程学院,河北 石家庄 050043
- 2. 石家庄铁道大学 机械工程学院,河北 石家庄 050043
- 3. 石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,河北石家庄 050043
- 4. 石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,河北石家庄 050043;石家庄铁道大学 机械工程学院,河北 石家庄 050043;石家庄铁道大学河北省大型工程机械装备制造协同创新中心,河北石家庄 050043
- 折叠
摘要
及时准确地检测路面裂缝,对于延长道路服役寿命、确保交通出行安全至关重要.针对现有方法在裂缝浅弱或背景噪声干扰下存在的分割断裂和误判漏检问题,以裂缝完整性为导向,提出了一种基于轻量化大核与低信息损耗的路面裂缝检测方法,可在复杂环境下实现路面裂缝的高效精准分割.首先,设计多尺寸大卷积核融合的特征编码结构,通过构筑多尺度的全局感受野,对路面大跨度裂缝的空间相关性进行充分建模.进而,提出正反注意力互补的重要性池化、训练-推理解耦的多支路解码、强调漏检惩罚的分层深监督等创新机制,全面缓解裂缝细节信息在编码与解码过程中出现的过度损耗,增强模型对裂缝细微特征的捕获和解析能力.同时,联合使用结构重参数化、部分卷积和深度卷积等策略,在确保高裕度特征表示的同时有效控制参数量,实现大核结构的超轻量化和推理模型的高效化.结果表明:该方法能够在复杂背景下实现路面裂缝的准确检测,所分割裂缝具有高完整性和低漏检率,同时具备轻量级的参数量和实时性的运算速度.与现有主流模型相比,所提方法的常规版本能够以3.67× 106的参数量取得77.17％的裂缝交并比、87.11％的Fi分数和87.41％的查全率,微缩版本则仅用0.29×106的超低参数量即可取得75.23％的裂缝交并比、85.63％的F分数和87.12％的查全率,其在将连通组件数偏差控制在极低水平的同时具备良好的路况适应性和噪声鲁棒性,在路面裂缝检测任务上有明显优势,能够为路面病害的监测、评估与修复提供有力支持.
Abstract
The timely and accurate detection of pavement cracks is crucial for extending the road service life and ensuring safe traffic.In response to the problems of segmentation discontinuity and false detection caused by weak cracks and background noise interference in existing methods,this study focuses on crack completeness and proposes a pavement crack detection method based on lightweight large kernels and low information waste to achieve efficient and accurate crack segmentation in a complex environment.First,a feature encoding structure based on multi-size large-kernel fusion was designed,which constructed a multiscale global receptive field to fully capture long-distance spatial dependencies for large-span pavement cracks.Second,innovative mechanisms such as importance-based pooling with forward-reverse complementary attention,diverse-branch decoding with training-inference decoupling,and hierarchical deep supervision with a reinforced misdetection penalty were constructed.These mechanisms aimed at alleviating the excessive information waste of crack details during both the encoding and decoding processes and improving the ability of the proposed model to capture and interpret fine-grained crack features.Further,this study combined structural reparameterization,partial convolution,and depthwise convolution to ensure effective control over the parameters while maintaining a highly expressive feature representation.Meanwhile,this strategy achieved an ultra-lightweight large-kernel structure and efficient model inference.The experimental results demonstrate that the proposed method can accurately segment pavement cracks in complex backgrounds with high crack completeness and low misdetection rates.In addition,it exhibits a lightweight parameter magnitude and real-time computational speed.The standard version of the proposed method can achieve an intersection-over-union rate of 77.17％,an F score of 87.11％,and a recall rate of 87.41％for cracks using 3.67× 106 parameters.The mini-version achieves an intersection-over-union rate of 75.23％,an Fi score of 85.63％,and a recall rate of 87.12％for cracks using only ultra-light 0.29X106 parameters.Moreover,the proposed method demonstrates good adaptation to road conditions and high robustness to background noise while controlling the quantity deviation of the connected components at an extremely low level.Compared with the current mainstream methods,the proposed method exhibits a significant advantage in pavement crack detection tasks and can strongly support the monitoring,evaluation,and repair of pavement de-fects.
关键词
路面工程/裂缝检测/完整性分割/大尺寸卷积核/超轻量级参数/结构重参数化Key words
pavement engineering/crack detection/complete segmentation/large-size convolution kernel/ultra-lightweight parameters/structural reparameterization引用本文复制引用
出版年
2024
NETL
NSTL
万方数据