A YOLO-based network with dynamic feature pyramid for multi-scale bridge surface defect detectionOA

Bridge surface defect detection is critical for guaranteeing transportation safety and maintenance planning.With the innovation of image acquisition technology and computer vision science,deep learning(DL)methods have revolutionized the ability to identify specific surface defects on bridges promptly.Bridge surface defect often exhibits different postures,appearing at different positions and scales in the image.However,the existing models need to be more satisfactory in detecting those surface defects at different scales,making them more challenging to detect and localize accurately by the models.To address this challenge,we proposed a more streamlined model,you only look once for bridge surface defect(YOLO-BSD),significantly improving the model performance on defect detection.We thoroughly evaluated our detection model’s effectiveness using a self-made dataset,and benchmarked its performance against other leading-edge models in the field.The experiments and results indicated that our proposed YOLO-BSD model surpasses current state-of-the-art(SOTA)algorithms in accuracy and operational efficiency,achieving a mean average precision(MAP)score of 0.892.Ablation experiments further demonstrated the effectiveness of our improvements,resulting in a total 6.6 increase in MAP.Our proposed model demonstrated superior detection capabilities and excelled in processing speed,achieving an impressive inference rate of 213(1000/4.7 ms)FPS,with only 10.6 G FLOPs and 12 MB parameters.