A Novel Network Fusing Transformer and CNN for Road Crack Segmentation

Abstract

Surface cracks pose a significant risk to the long-term strength and security of roads, frequently resulting in safety issues. As a result, crack identification is of utmost importance in the examination and assessment of roads during maintenance. However, existing object detection methods either exhibit low accuracy in crack detection or have large model sizes, leading to unsatisfactory results. This research presents a novel network that aims to achieve a balance between accuracy and model size. The network strives to enhance detection precision while reducing the size requirements of the model. The method entails integrating transformer blocks into convolutional neural networks to augment their receptive fields. Furthermore, a novel AFM model is introduced to enhance the neural network’s capacity to acquire and integrate both high-level and low-level feature maps. Finally, a lightweight decoding block (SMLP) is utilized to consolidate data from several levels, efficiently using lower-level feature map information for thorough crack detection. The proposed network was assessed using two publicly available crack datasets, consistently outperforming other approaches in several evaluations. Notably, it attained the highest IoU scores on the DeepCrack and CrackTree200 datasets, reaching impressive values of 0.7812 and 0.7463, respectively. Overall, the proposed network outperformed other advanced networks in fracture segmentation accuracy. It was also highly generalizable across crack widths, proving its versatility.

Keywords

• Deep learning
• crack segmentation
• transformer
• feature fusion enhancement module