Asphalt Pavement Crack Detection Based on Improved YOLOv8

To address the issues of low efficiency and high cost in asphalt pavement crack detection, a low-cost solution using smartphones is proposed. First, to improve the overall detection performance of the model, a FreqSpatialFusion module is designed and implemented to extract both spatial and frequency domain features from images. Additionally, the original module is replaced with a WIoUv3 loss function and a dual-channel Spatial Pyramid Pooling Faster Cross Stage Partial Channel (SPPFCSPC) layer to enhance the model's recognition ability. Secondly, to reduce model redundancy, a Layer Adaptive Magnitude Based Pruning (LAMP) technique is employed for channel pruning, thereby reducing the model's size and computational load. Finally, the improved model, YOLOv8n-Crack, is deployed on smartphones to enable low-cost crack detection. Experimental results show that the improved model achieves a 4.2% increase in mAP@50 on the self-built Crack-5514 dataset, while reducing the model size and computational load by 36% and 8.6%, respectively. This approach enables real-time crack detection, significantly improving efficiency and reducing costs.