Structural crack detection using deep convolutional neural network / Raza Ali

Raza , Ali (2022) Structural crack detection using deep convolutional neural network / Raza Ali. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (221Kb)
	PDF (Thesis PhD) Download (4021Kb)

Abstract

Convolutional Neural Networks (CNN) have immense potential to solve a broad range of computer vision (CV) problems. It has achieved encouraging results in numerous applications in engineering, medical, and other research fields. Thanks to the advancement in hardware, data collection procedures, and efficient algorithms. These innovations have changed the way how specific problems are solved as compared to conventional methods. In this work, CNN is implemented for civil structural crack detection. Cracks are significant indicators for the evaluation of the structural health and monitoring process. However, manual crack detection is a time-consuming and challenging task due to large areas, complex structures, and safety risks. Deep learning (DL) has emerged as an effective technique to automate the crack detection and identification process. For balanced data, existing DL models attempt to segment both crack pixels and non-crack pixels equally. However, due to the highly imbalanced ratio between crack pixels and non-crack pixels, the pixel-wise loss is dominantly guided by the non-crack region and has relatively little influence from the crack region. This leads to the low segmentation accuracy for crack pixels. To address the imbalance problem, this work proposes a local weighting factor with a difference transform map to remove the network biasness and accurately predict the sensitive pixels. Further, a deep fully CNN called crack segmentation network (CSN) is implemented for crack pixel segmentation. The CSN is an encoder-decoder architecture with four convolutional blocks in each section. Each convolutional block has residual connections with a different number of filters in each convolutional operation that segments the crack pixels and non-crack pixels with unbiased probabilities. Furthermore, the crack indicators are assessed through the implementation of a pixel connection technique that measures the crack characteristics (length, width, and area) and determines the crack orientation (vertical, horizontal, and diagonal). For performance evaluation, a new Multi Structure Crack Image (MSCI) dataset is built to train the proposed method which achieved 98.60% crack pixel accuracy, 98.35% non-crack pixel accuracy, and 98.48% average accuracy, respectively. In addition, the training time for 10 epochs has dramatically decreased and the experimental results show that the proposed CSN architecture has better crack pixel segmentation accuracy than FCN, U-Net, SegNet, and DeepLabv3+ architectures. Similarly, the proposed local weighting factor and difference transform map (LWF-DTM) has significantly reduced the wrong predictions, minimized the effect of an imbalanced pixel ratio, and outperformed the Cross-Entropy, Weighted Cross-Entropy, Dice, Tversky, and Focal loss function.

Actions (For repository staff only : Login required)