GDAN-Deeplabv3+: a crack detection method fusing attention mechanism and Deeplabv3+

Crack detection is a typical structural application in computer vision. In many scenarios, such as road cracks, cracks often exhibit weak continuity and low contrast, posing significant challenges for crack detection using low-level features. This paper proposes a dual-attention deep neural network for accurately classifying road cracks and locating their positions. The innovative detection model, GDAN-Deeplabv3+, incorporates dual attention modules into the Deeplabv3+ architecture to attain improved segmentation accuracy. Firstly, the ASPP module within the Deeplabv3+ framework is proficient in capturing multi-scale contextual data and defining clear object boundaries. Then, the dual-parallel attention modules explicitly model local features in both spatial and channel dimensions, enabling better focus on relevant features. GDAN-Deeplabv3+ was experimentally validated on the Crack500 and DeepCrack datasets. GDAN-Deeplabv3+ based on the ResNet-101 backbone achieved mean Intersection over Union (mIoU) scores of 83.45% and 91.18% on the two datasets, respectively. Comparative studies indicate that incorporating dual-parallel attention modules and the Deeplabv3+ framework improves segmentation performance. The proposed hybrid model enhances segmentation performance in crack detection and has been demonstrated to outperform other classical methods.