Crack Detection as a Weakly-Supervised Problem: Towards Achieving Less Annotation-Intensive Crack Detectors

Abstract

Automatic crack detection is a critical task that has the potential to drastically reduce labor-intensive building and road inspections currently being done manually. Recent studies in this field have significantly improved the detection accuracy. However, the methods often heavily rely on costly annotation processes. In addition, to handle a wide variety of target domains, new batches of annotations are usually required for each new environment. This makes the data annotation cost a significant bottleneck when deploying crack detection systems in real life. To resolve this issue, we formulate the crack detection problem as a weakly-supervised problem and propose a two-branched framework. By combining predictions of a supervised model trained on low quality annotations with predictions based on pixel brightness, our framework is less affected by the annotation quality. Experimental results show that the proposed framework retains high detection accuracy even when provided with low quality annotations. Implementation of the proposed framework is publicly available at https://github. com/hitachi-rd-cv/weakly-sup-crackdet.

What's Cracking? A Review and Analysis of Deep Learning Methods for Structural Crack Segmentation, Detection and Quantification

“…The acquisition of data for this approach can pose a problem for large-scale applications. Depending on the level of detail required in the labels, it can take up to several minutes to label a single image [83]. This is a domain-wide issue with ML for computer vision often requiring large amounts of labeled data to train effectively.…”

“…Using those synthetic, binary-segmentation labels a fully convolutional segmentation network is trained to predict cracks. The work in [83] performs weakly supervised crack segmentation by reducing the quality of labels through a dilation process and add an additional component that can be added to any crack segmentation approach to increase performance. This second component is not DL based but rather exploits the crack characteristic that they are darker than the usual structure background.…”

“…2) Label Quality: Labeling crack data can be a difficult, time-consuming and tedious process that is subjective to human error [3]. Depending on the type of label, thorough labeling of a single image can take up to several minutes [83]. This can be an expensive process as a specialist might have a high hourly charge and using crowd-sourcing to label images such as in ImageNet may exceed the initial budget.…”

“…with T being the total number of thresholds used. Other metrics that have also been used is the area under the precision-recall curve reported as the average precision [33], [68], [120], Macro-F1 [83], as well as the Matthews correlation coefficient [18].…”

“…Here, similar to the previously mentioned dataset issues, it may be beneficial to create a standardized dataset for those types of learning. Current approaches have authors changing the label quality [81], [83] of given datasets manually, which may not be the most optimal way going forward. A unified dataset for those types of learning can accelerate research and provide better comparability of given algorithms.…”