FissureNet: A Concrete Bridge Segmentation Method Based on High-Resolution Images from UAV

Accurate semantic segmentation of concrete bridge cracks is crucial for road health monitoring. Currently, crack images are primarily obtained through manual field inspections. As an automated robotic technology, Unmanned aerial vehicles (UVAs) can provide high-resolution images, offering a new, efficient, and cost-effective method for rapid semantic segmentation of cracks. However, the fine local structures, elongated continuous global forms, and the imbalance between crack pixels and background pixels make it challenging for traditional convolutional networks to perform this task effectively. To address these challenges, we have introduced FissureNet, a network based on an enhanced DeepLabv3+ model. Initially, the network utilizes a lightweight architecture to replace the original backbone, thereby reducing model complexity. Subsequently, inspired by the biological movement of a snake swaying side to side, we introduced dynamic strip convolution (DPconv), where the convolution kernels can twist like a snake to better fit our target crack objects. Enriched with examples of bio-inspired applications of biomimetics in the computer field . All atrous convolution branches in the original atrous spatial pyramid pooling (ASPP) module have been replaced with our newly proposed DPconv. This convolution adaptively focuses on small local structures. Moreover, we have incorporenabling more precise extraction of crack features.ated a multi-perspective feature extraction strategy, introducing a vision transformer (VIT) to capture global features. This strategy is combined with traditional convolution to form a hybrid structure that intensifies focus on features from both global and local perspectives, enhancing the continuity of narrow crack segmentation. Lastly, an attention gata mechanism introduced in the decoding phase improves feature extraction near crack edges and significantly reduces the disparity between crack and background pixel counts. Empirical validation on the SDNET2018 dataset has demonstrated that FissureNet outperforms several classical methods in terms of accuracy and continuity in segmenting concrete bridge cracks.