YOLOv8-DAM: Pavement Disease Study Based on Improved YOLOv8

To address the challenges of complex background interference, poor small target detection, and excessive model parameters in pavement disease detection, we propose YOLOv8-DAM, an improved road disease detection algorithm based on YOLOv8n. Three key innovations are introduced: First, we design a Dynamic Weight Pyramid Network (DWPN) module that integrates the SE mechanism to weight input features, effectively distinguishing target features from the background. Second, we develop the ADown module, combining multiple pooling operations with parallel structures to enhance small target detection accuracy. Third, we redesign the Detect_MSFD detection head with shared convolutional layers, reducing parameter redundancy by 31.7% while improving inference speed by 26.2% in FPS. Experimental results on the RDD2022 dataset show mAP@0.5 of 54.7% and mAP@0.5:0.95 of 31.2%, representing 3.2% and 1.3% improvements over the baseline model, respectively. Additional validation on the public SVRDD dataset demonstrates further gains of 3.5% in mAP@0.5 and 2% in mAP@0.5:0.95, confirming the algorithm’s strong generalization capability.