No abstract

Conventional surface crack segmentation requires images manually labelled by a trained expert. It is a challenging task as cracks can vary in orientation and size, with some parts of cracks only being one pixel wide. Further, available training data for crack segmentation is sparse. In this work we propose to automate this annotation task, by introducing a fully convolutional U-Net based architecture for semantic segmentation of surface cracks which allows for the use of small datasets through a patch based training process. Our proposed configuration makes use of residual connections inside the convolutional blocks as well as including an attention based gating mechanism between the encoder and decoder section of this architecture, which only propagates relevant activations further. Using our proposed architecture we achieve new state of the art results in two different crack datasets, outperforming the previous best results in two metrics each.

Continual use, as well as aging, allows cracks to develop on concrete surfaces. These cracks are early indications of surface degradation. Therefore, regular inspection of surfaces is an important step in preventive maintenance, allowing reactive measures in a timely manner when cracks may impair the integrity of a structure. Automating parts of this inspection process provides the potential for improved performance and more efficient resource usage, as these inspections are usually carried out manually by trained inspectors. In this work we propose a Fully Convolutional, U-Net based, Neural Network architecture to automatically segment cracks. Conventional pooling operations in Convolutional Neural Networks are static operations that reduce the spatial size of an input, which may lead to loss of information as features are discarded. In this work we introduce and incorporate a novel pooling function into our architecture, Gated Scale Pooling. This operation aims to retain features from multiple scales as well as adapt proactively to the feature map being pooled. Training and testing of our network architecture is conducted on three different public surface crack datasets. It is shown that employing Gated Scale Pooling instead of Max Pooling achieves superior results. Furthermore, our experiments also indicate strongly competitive results when compared with other crack segmentation techniques.

No abstract

Surface cracks are a common sight on public infrastructure nowadays. Recent work has been addressing this problem by supporting structural maintenance measures using machine learning methods. Those methods are used to segment surface cracks from their background, making them easier to localize. However, a common issue is that to create a wellfunctioning algorithm, the training data needs to have detailed annotations of pixels that belong to cracks. Our work proposes a weakly supervised approach that leverages a CNN classifier in a novel way to create surface crack pseudo labels. First, we use the classifier to create a rough crack localization map by using its class activation maps and a patch based classification approach and fuse this with a thresholding based approach to segment the mostly darker crack pixels. The classifier assists in suppressing noise from the background regions, which commonly are incorrectly highlighted as cracks by standard thresholding methods. Then, the pseudo labels can be used in an end-toend approach when training a standard CNN for surface crack segmentation. Our method is shown to yield sufficiently accurate pseudo labels. Those labels, incorporated into segmentation CNN training using multiple recent crack segmentation architectures, achieve comparable performance to fully supervised methods on four popular crack segmentation datasets.