基于矢量量化变分自编码器的混凝土表观裂缝检测算法

Detection Algorithm of Concrete Structural Apparent Cracks Based on VQ-VAE-2

摘要:

提出了一种基于第2代矢量量化变分自编码器（VQ?VAE?2）的自监督混凝土表观裂缝检测算法，可以在缺少裂缝样本的条件下实现高效检测。以重建误差为检测指标，利用无裂缝图片训练VQ?VAE?2，使其在重建裂缝图片时产生更大的重建误差；在计算重建误差时将原图和重建图片均分割成若干图块，取对应图块间重建误差最大值作为图片的重建误差，以增大2类图片的重建误差差异。结果表明，该算法的精确率为0.954，召回率为0.959，准确率为0.956，F1分数为0.957。在无裂缝样本作为训练集的情况下，该算法能较好地完成混凝土表观裂缝检测任务。

关键词:桥梁工程;混凝土表观裂缝检测;深度学习;变分自编码器;异常检测

Abstract:

In this paper， we propose a self-supervised algorithm based on VQ-VAE-2 for the automated detection of concrete structural apparent cracks. The algorithm demonstrates the capability to effectively detect cracks without available crack samples. By taking the reconstruction error as detection index， VQ-VAE-2 is trained on crack-free images. When applied to images with cracks， VQ-VAE-2 produces higher reconstruction errors. The original and reconstructed images are partitioned into blocks for calculating the reconstruction error. The maximum value of the reconstruction error between corresponding blocks is taken as the reconstruction error of the image. This approach increases the difference in reconstruction error between the two types of images. The results show that the algorithm achieves a precision of 0.954， a recall of 0.959， an accuracy of 0.956， and an F1 score of 0.957. These results indicate that the algorithm can effectively detect concrete structural apparent cracks even without crack samples in the training set.

Key words:bridge engineering;concrete structural apparent crack detection;deep learning;variational autoencoders;anomaly detection