YOLOv11-AESNet: Efficient Real-Time Crack Segmentation Using Deep Learning Techniques

To address the challenges of variable crack morphology and insufficient real-time performance in infrastructure crack detection, this paper proposes a lightweight real-time crack instance segmentation network, YOLOv11-AES, based on an improved YOLOv11. First, a deformable kernel convolution module is introduced to dynamically adjust the offset and scale factor of convolutional kernels, enhancing the model’s ability to extract features from irregular cracks. Second, an efficient hierarchical attention mechanism is adopted to improve focus on low-contrast cracks through cross-channel and spatial collaboration. Finally, a decoupled segmentation head is proposed to optimize mask generation accuracy while reducing computational overhead. Experiments on a self-built dataset show that YOLOv11-AES achieves a detection accuracy of 96.8% ($\mathrm{mAP}_{50}{ }^{\text {Box }}$), a segmentation quality of $90.8 \%\left(\mathrm{mAP}_{50}{ }^{\text {Mask }}\right)$, and an inference speed of 63.7 frames per second (FPS), with significantly fewer parameters ($\mathbf{2. 5 M}$) and lower computational complexity (9.4G FLOPs) compared to mainstream instance segmentation models (e.g., YOLOv8-seg, SOLOv2). This study provides an efficient and reliable solution for intelligent crack detection in infrastructure, balancing high accuracy and real-time performance.