CSegNet：A Crack Segmentation Network Combining CNN and Transformer

Abstract Detecting cracks from images plays a crucial role in road maintenance. Road cracks exhibit significant diversity and complexity in terms of shape, size, texture, and road images may contain various noises and interferences such as lighting variations, shadows, and different appearances due to varying perspectives and scales. To address these challenges, we constructed a comprehensive dataset called the Comprehensive Road Crack Dataset (CRCrack Dataset), which encompasses various crack characteristics. In this study, we propose a road crack segmentation network called CSegNet (Crack Segmentation Network), which combines convolutional neural networks (CNNs) and Transformers. The network adopts an encoder-decoder framework, like DeepLab V3+. In the encoder, leveraging the flexibility of Transformers in modeling long-term dependencies and the ability of CNNs to capture local contextual information through local receptive fields, weight sharing, and spatial subsampling, we design a ResNeXTR (ResNeXt-Transformer) feature extraction module as the backbone network to enhance the feature extraction capability for road crack images. To reduce the computational cost in self-attention computation of transformer, we introduce an average pooling layer to downsample the dimensions of the encoded features. In the decoder, to focus on the key information of road cracks under diverse environmental conditions and interferences, we combine the Efficient Channel Attention Module (ECAM) and the Spatial Attention Module (SAM) to design an Efficient Convolutional Block Attention Module (ECBAM) attention module to further optimize feature representation. Additionally, we employ the ReLU activation function, SGD gradient descent, and a hybrid loss function of Binary Cross Entropy with Logits to accelerate convergence speed and improve segmentation accuracy. Through comparative experiments on the CRCrack dataset, the results demonstrate that our proposed method outperforms classic networks such as U-Net and DeepLab V3 + in terms of IoU, Dice, and AUROC evaluation metrics. It exhibits good adaptability to ground crack images from different sources, providing a basis for estimating the degree of road damage.