Deep crack segmentation: a semi-supervised approach with coordinate attention and adaptive loss

Abstract As infrastructure ages, crack detection and segmentation become increasingly important in maintenance and management. However, traditional fully-supervised learning methods rely on a large amount of labeled data, which is often difficult to obtain in practical applications, limiting the generalization ability of the model. To address this problem, we propose a new semi-supervised crack segmentation method-Semi-CSN, which aims to improve model performance by leveraging unlabeled data. The road crack segmentation network proposed in this study achieves accurate crack segmentation in complex scenes through the co-design of multi-scale feature extraction, multi-scale feature fusion (MFF)_ and semi-supervised learning (SSL). The model adopts the pre-trained ResNet-50 as the backbone network to generate feature maps at four different scales, and processes the deep feature maps through the improved atrous spatial pyramid pooling (ASPP) structure to effectively capture the slender geometric features of the cracks. The feature maps at different scales are integrated by an attention-based feature fusion module to enhance the extraction of global and local information. In the SSL framework, we design a new joint label-pseudo-label training strategy to co-optimize labeled and unlabeled data to enhance the performance of crack detection. Specifically, the labeled data are used to compute the supervised loss to ensure the effective training, while the unlabeled data are processed with weak and strong enhancement and generate high-quality pseudo-labels based on the dynamic thresholding strategy, which effectively improves the efficiency of unlabeled data utilization. In addition, a new loss function is designed to adaptively increase the weight of the sparser regions in the crack image. Experiments conducted on two publicly available datasets show that our method significantly improves segmentation accuracy and outperforms other baseline networks in the crack segmentation task.