Pixel-level crack delineation in images with convolutional feature fusion

Cracks in civil structures are important signs of structural degradation and may even indicate the inception of catastrophic failure. Image-based crack detection has been attempted in research communities that bear the potential of replacing human-based inspection. Among many methodologies, deep learning-based cracks detection is actively explored in recent years. However, how to automatically extract cracks quickly and accurately at a pixel level, that is, crack delineation (including both detection and segmentation), is a challenging issue. This article proposes a convolutional neural network-based framework that automates this task through convolutional feature fusion and pixel-level classification. The resulting network architecture with an empirically optimal fusion strategy, termed the crack delineation network, is trained and tested based on a concrete crack image database. The results show that the proposed framework can delineate cracks accurately and rapidly in images towards a fully autonomous machine vision approach to structural crack detection.