When Transfer Learning Meets Dictionary Learning: A New Hybrid Method for Fast and Automatic Detection of Cracks on Concrete Surfaces

Abstract

Cracks in civil structures are important signs of structural degradation and may indicate the inception of catastrophic failure. However, most of studies that have employed deep learning models for automatic crack detection are limited to high computational demand and require a large amount of labeled data. Long training time is not friendly to model update, and large amount of training data is usually unavailable in real applications. To bridge this gap, the innovation of this study lies in developing a hybrid method that comprises transfer learning (TL) and low-rank dictionary learning (LRDL) for fast crack detection on concrete surfaces. Benefiting from the availability of preextracted features in TL and a limited number of parameters in LRDL, the training time can be significantly minimized without GPU acceleration. Experimental results showed that the time for training a dictionary only takes 25.33 s. Moreover, this new hybrid method reduces the demand for labeled data during training. It achieved an accuracy of 99.68% with only 20% labeled data. Three large-scale images captured under varying conditions (e.g., uneven lighting conditions and very thin cracks) were further used to assess the crack detection performance. These advantages help to implement the proposed TL-LRDL method on resource-limited computers, such as battery-powered UAVs, UGVs, and scarce processing capability of AR headsets.