Efficient attention-based deep encoder and decoder for automatic crack segmentation

Abstract

Recently, crack segmentation studies have been investigated using deep convolutional neural networks. However, significant deficiencies remain in the preparation of ground truth data, consideration of complex scenes, development of an object-specific network for crack segmentation, and use of an evaluation method, among other issues. In this paper, a novel semantic transformer representation network (STRNet) is developed for crack segmentation at the pixel level in complex scenes in a real-time manner. STRNet is composed of a squeeze and excitation attention-based encoder, a multi head attention-based decoder, coarse upsampling, a focal-Tversky loss function, and a learnable swish activation function to design the network concisely by keeping its fast-processing speed. A method for evaluating the level of complexity of image scenes was also proposed. The proposed network is trained with 1203 images with further extensive synthesis-based augmentation, and it is investigated with 545 testing images (1280 × 720, 1024 × 512); it achieves 91.7%, 92.7%, 92.2%, and 92.6% in terms of precision, recall, F1 score, and mIoU (mean intersection over union), respectively. Its performance is compared with those of recently developed advanced networks (Attention U-net, CrackSegNet, Deeplab V3+, FPHBN, and Unet++), with STRNet showing the best performance in the evaluation metrics-it achieves the fastest processing at 49.2 frames per second.

Sparse‐sensing and superpixel‐based segmentation model for concrete cracks

“…In the experiments, six recently published networks, including U‐Net (Z. Liu et al., 2019), DeepCrack (Y. Liu et al., 2019), U‐Net++ (Zhou et al., 2019), Attention U‐Net (König et al., 2019), DeepLabv3+ (L.‐C. Chen et al., 2018; Ji et al., 2020), and semantic transformer representation network (STRNet; Kang & Cha, 2021), were selected for their state‐of‐the‐art performances in the crack segmentation. All the networks as listed in Table 1 were implemented using the same dataset and environment.…”

“…The added convolution in the attention gate sacrificed the width and depth of the network, which may interfere with the identification performance. Kang and Cha (2021) improved the encoder and decoder with squeeze-excitation-based and attention-based modules hoping to make the semantic representation of the network trainable. However, the matrix multiply in the attention decoder breaks the translation invariance of CNN, which probably degrades the performance on the images that are not the same size as the training images.…”

Autogenous self‐healing of concrete: Experimental design and test methodsA review

“…However, the detection based on bounding-boxes is too coarse to quantify defects such as the crack length and width, which requires pixel-level crack segmentation. 72,73 For instance, Kang et al 74 modified a R-CNN algorithm to allow for a crack segmentation into pixels and further measurement of the crack thickness and length by pixel analysis. The latter authors achieved an accuracy of 93%.…”

A New Method for Automated Monitoring of Road Pavement Aging Conditions Based on Recurrent Neural Network

“…Thus the approach performs well on multiple sample sets. Kang et al [18], [47], [48] perform crack segmentation in complex environments and different lighting conditions by integrating three independent computer vision algorithms and developed a new encoder with an attention module. Choi et al [49] propose a real-time crack segmentation DL architecture, referred to as SDDNet-V1, which can greatly improve the time efficiency and identify relatively vague cracks.…”