Automatic Concrete Defect Identification by Silencing Features of Deep Neural Network

Abstract

An autonomous concrete crack inspection system is necessary for preventing hazardous incidents arising from deteriorated concrete surfaces. In this thesis, we represent a concrete crack detection framework to aid the process of automated inspection.Deep neural networks highly suffer from the gradient vanishing problem [1]. The effect of gradient vanishing problem is very prominent on class imbalanced data-setssuch as crack detection. In this work, a deep neural architecture is proposed foralleviating the effect gradient vanishing problem. Furthermore, A feature silencingmodule is incorporated in the crack detection framework, for eliminating unnecessaryfeature maps from the network. This module reduces the computational costs of deepneural networks. The overall performance of the network significantly improves asa result. Experimental results support the benefit of incorporating feature silencingwithin a convolutional neural network architecture for improving the network’s robustness, sensitivity, and specificity. An added benefit of the proposed architecture isits ability to accommodate for the trade-off between specificity (positive class detection accuracy) and sensitivity (negative class detection accuracy) with respect to thetarget application. Furthermore, the proposed framework achieves a high precisionrate and processing time than crack detection architectures present in literature.