基于改进YOLO11的高速路面裂缝分割算法OA

To address the challenges including the high missed detection rates,insufficient boundary localization accuracy,and poor adaptability to complex environments in the high-speed pavement crack segmentation tasks under the IoT edge environment,an improved YOLO11 segmentation model is proposed.This model integrates lightweight multi-scale feature enhancement and a Polarized Self-Attention(PSA)mechanism to achieve a synergistic optimization between high precision and low latency.A lightweight design of the Atrous Spatial Pyramid Pooling(ASPP)module is introduced,which employs parallel atrous convolutions with different dilation rates and global average pooling to enhance multi-scale contextual information extraction capability while reducing computational complexity.Meanwhile,an optimized PSA mechanism is incorporated to jointly recalibrate features across channel and spatial dimensions,effectively suppressing complex background interference and improving crack edge recognition accuracy.This mechanism enables edge nodes to transmit only sparse features filtered by attention weights,thereby reducing data transmission overhead and accommodating multi-node collaborative perception in the IoT environment.Experiments on the self-built crack dataset from the G85 highway demonstrate that the improved model exhibits strong robustness under complex backgrounds and low-contrast conditions.Its overall performance has improved by 5.62%compared to the YOLO11n baseline,and it has also significantly outperformed the state-of-the-art methods.The proposed model provides a high-precision,lightweight segmentation solution for intelligent highway pavement maintenance.