Weakly-Supervised Surface Crack Segmentation by Generating Pseudo-Labels Using Localization With a Classifier and Thresholding

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.

CrackCLF: Automatic Pavement Crack Detection Based on Closed-Loop Feedback

“…Guo et al [36] proposed BARNet method, which includes three modules: Edge Adaptation Module, Base Predictor Module and Refinement Module. Knig [37] proposed using the Weakly-supervised method to segmentation crack image, which can obtain a better accuracy than others. Sun et al [38] proposed a DMA (DeepLab With Multi-Scale Attention) method based on DeepLab, which integrates the ASPP(Atrous Spatial Pyramid Pooling) and attention mechanism.…”

CAC: Confidence-Aware Co-Training for Weakly Supervised Crack Segmentation

“…These pseudo-labels are then used to train the crack segmentation model. König et al [12] proposed a WSCS method that employed location with a classifier and threshold segmentation to generate crack pixel-level pseudo-labels. These pseudo-labels serve as training data for the crack segmentation model.…”

“…The crack pseudo-labels are generated using two distinct methods for comparison. The first method employs location information with the crack classifier and threshold segmentation as proposed by König et al [12]. The second method utilizes CRF post-processing to refine the crack CAMs, following the approach by Dong et al [13].…”

“…To address this issue, weakly supervised crack segmentation (WSCS) attracts increasing attention [10]. The weakly supervised segmentation only needs coarse-grained information such as the bounding boxes [11], and patch-level labels [12][13][14]. Among them, the patch level labels greatly reduce annotation difficulty by only giving the category information of image patches, which are thus widely researched in the existing literature of WSCS.…”

“…The majority of existing WSCS methods primarily concentrate on improving the accuracy of pixel pseudo-label generation. König et al [12] introduced a WSCS method that produces high-quality pixel pseudo-labels by integrating class activation maps (CAMs) [15] with classifier localization and threshold segmentation. Similarly, Dong et al [13] proposed a WSCS model rooted in patch-based techniques.…”