Robust Pixel-Level Crack Detection Using Deep Fully Convolutional Neural Networks

摘要

This paper introduces the idea of using deep fully convolutional neural networks for pixel-level defect detection in concrete infrastructure systems. Although coarse patch-level deep learning crack detection models abound in the literature and have shown promise, the coarse level of detail provided, together with the requirement for fixed-size input images, significantly detract from their applicability and usefulness for refined damage analysis. The deep fully convolutional model for crack detection introduced in this paper (CrackPix) leverages well-known image classification architectures for dense predictions by transforming their fully connected layers into convolutional filters. A transposed convolution layer is then used to upsample and resize the resulting prediction heatmap to the size of the input images, thus providing pixel-level predictions. To develop and train these models, a concrete crack image data set was collected and carefully annotated at the pixel level and was then used to train the model. Sensitivity analysis showed that CrackPix was capable of correctly detecting over 92% of crack pixels and 99.9% of noncrack pixels in the validation set. The model performance was then compared against a state-of-the-art patchwise model, as well as traditional edge detection and adaptive thresholding alternatives, and its advantages were illustrated. The success of CrackPix, which enables the quantification of crack characteristics (e.g., width and length) in concrete structures, provides a key step toward automated inspection and quality assurance for infrastructure in future smart cities.