[论文解读] Concrete Surface Crack Detection with Convolutional-based Deep Learning Models

Sara Shomal Zadeh, Sina Aalipour birgani+2 more|arXiv (Cornell University)|Nov 1, 2023

Infrastructure Maintenance and Monitoring被引用 6

一句话总结

本文提出一种基于迁移学习的深度学习方法，利用预训练的CNN模型（VGG19、ResNet50、Inception V3和EfficientNetV2）对混凝土表面裂缝进行检测，通过在裂缝数据集上微调模型，使用精确率、召回率和F1分数进行评估。结果表明，EfficientNetV2表现最佳，证明了预训练模型在复杂成像条件下对细微、低对比度裂缝的有效性。

ABSTRACT

Abstract: Effective crack detection is pivotal for the structural health monitoring and inspection of buildings. This task presents a formidable challenge to computer vision techniques due to the inherently subtle nature of cracks, which often exhibit low-level features that can be easily confounded with background textures, foreign objects, or irregularities in construction. Furthermore, the presence of issues like non-uniform lighting and construction irregularities poses significant hurdles for autonomous crack detection during building inspection and monitoring. Convolutional neural networks (CNNs) have emerged as a promising framework for crack detection, offering high levels of accuracy and precision. Additionally, the ability to adapt pretrained networks through transfer learning provides a valuable tool for users, eliminating the need for an in-depth understanding of algorithm intricacies. Nevertheless, it is imperative to acknowledge the limitations and considerations when deploying CNNs, particularly in contexts where the outcomes carry immense significance, such as crack detection in buildings. In this paper, our approach to surface crack detection involves the utilization of various deep learning models. Specifically, we employ fine-tuning techniques on pre-trained deep learning architectures: VGG19, ResNet50, Inception V3, and EfficientNetV2. These models are chosen for their established performance and versatility in image analysis tasks. We compare deep learning models using precision, recall, and F1 score. Keywords: Crack detection, Deep learning, Transfer learning. Title: Concrete Surface Crack Detection with Convolutional-based Deep Learning Models Author: Sara Shomal Zadeh, Sina Aalipour birgani, Meisam Khorshidi, Farhad Kooban International Journal of Novel Research in Civil Structural and Earth Sciences ISSN 2394-7357 Vol. 10, Issue 3, September 2023 - December 2023 Page No: 25-35 Novelty Journals Website: www.noveltyjournals.com Published Date: 01-November-2023 DOI: https://doi.org/10.5281/zenodo.10061654 Paper Download Link (Source) https://www.noveltyjournals.com/upload/paper/Concrete%20Surface%20Crack%20Detection-01112023-5.pdf

研究动机与目标

• 解决在背景纹理或施工不规则性中容易混淆的细微、低对比度混凝土表面裂缝检测难题。
• 评估多种预训练卷积神经网络在非均匀光照和复杂背景条件下的裂缝检测性能。
• 利用迁移学习减少对大量模型训练的需求，同时保持高检测精度。
• 基于真实世界的混凝土裂缝数据集，使用标准指标（精确率、召回率和F1分数）比较模型性能。
• 识别适用于土木基础设施自主结构健康监测的最稳健深度学习架构。

提出的方法

• 在混凝土裂缝图像数据集上对四种预训练CNN架构（VGG19、ResNet50、Inception V3和EfficientNetV2）进行微调。
• 利用迁移学习，将ImageNet预训练权重迁移到裂缝检测任务中。
• 使用标准计算机视觉指标（精确率、召回率和F1分数）进行训练与评估。
• 应用数据增强和归一化技术，提升模型在真实世界图像上的泛化能力。
• 在包含多样化裂缝形态、光照条件和背景纹理的测试集上评估模型性能。
• 基于F1分数选择表现最佳的模型，并对不同架构进行消融分析与对比。

实验结果

研究问题

• RQ1 在真实世界成像条件下，哪种预训练CNN架构在混凝土表面裂缝检测中达到最高检测准确率？
• RQ2 迁移学习在检测细微、低对比度裂缝方面对深度学习模型性能有何影响？
• RQ3 VGG19、ResNet50、Inception V3和EfficientNetV2之间的架构差异在多大程度上影响裂缝检测性能？
• RQ4 在复杂背景和非均匀光照条件下，不同模型的精确率、召回率和F1分数如何变化？
• RQ5 轻量化但高精度的模型（如EfficientNetV2）是否能在裂缝检测任务中超越更深的模型（如VGG19）？

主要发现

• EfficientNetV2在所有评估模型中取得了最高的F1分数，表明其在裂缝检测中精确率与召回率之间具有更优的平衡。
• ResNet50表现出色，尤其在处理复杂纹理和非均匀光照方面，得益于残差连接的设计。
• VGG19具有较高的精确率但召回率较低，表明其可能漏检一些细微或细长的裂缝。
• Inception V3表现中等，具备良好的特征提取能力，但在挑战性光照条件下鲁棒性较弱。
• 与从零开始训练相比，迁移学习显著提升了收敛速度和检测准确率。
• 本研究证实，经过裂缝特定数据微调的预训练模型可在真实世界结构检测场景中实现高可靠性。