Vision-Based Semi-Supervised Learning Method for Concrete Crack Detection: Fid-Bau Portal

Vision-Based Semi-Supervised Learning Method for Concrete Crack Detection

Liu, Yiqing / Yeoh, Justin K. W.

Building defects inspection and maintenance needs to be carried out periodically. Cracking is one of the most common and essential defects that provides an indicator of the structural health of the building. Current inspection methods are carried out manually, and suffer from personnel safety problems which are both labor-intensive and time-consuming. In order to overcome the limitations of manual visual inspection, some image-based crack detection methods using computer vision and machine learning algorithms have been developed. With recent advances in deep learning techniques, convolution neural networks (CNNs) are gaining prominence as tools to detect cracks from images of the building. However, most current learning-based methods are implemented in a fully supervised manner which requires large amount of labeled data for training. To address issues of scalability, generalizability, and overfitting, large amounts of data need to be collected and labeled manually which is time-consuming and labor-intensive. In this paper, we explore the application of a semi-supervised learning method for crack detection to mitigate the above problem. In this method, a deep generative adversarial network (GAN) is trained as a classifier to identify the images with crack and without crack in the dataset. In order to validate the efficiency of the semi-supervised learning method, we tested the supervised learning method which applies transfer learning to convolution neural network for crack detection on the same dataset. The results showed that the semi-supervised learning method could achieve a similar accuracy of 98% with 62.5% labeled images compared with supervised learning method. The implication of our results suggest the ability to achieve a reasonably high detection accuracy with a low proportion of labeled images.