| To address practical engineering challenges in pipeline defect detection, including fisheye lens imaging distortion, uneven illumination, large variations in defect scale, and severe class imbalance, this paper proposes an improved pipeline detection model named Pipe-YOLO, using YOLOv11n as the baseline. First, to alleviate the long-tail distribution problem in pipeline defect data, a targeted offline data augmentation strategy is applied to the Pipe Dataset (V2). By simulating multi-dimensional imaging conditions, this strategy improves inter-class balance and enhances the model’s discriminative ability for minority defect categories. Second, to cope with geometric distortion and non-uniform illumination caused by fisheye imaging, a Pipeline Convolutional Block Attention Module (PCBAM) is designed. Through radial feature enhancement and cross-channel spatial collaboration, PCBAM improves the model’s feature representation capability in complex backgrounds. Then, to address the precise localization problem of multi-scale defects, an Adaptive Boundary Intersection over Union loss function (ABIoU) is introduced, which optimizes the detection of small defects by dynamically adjusting boundary weights. Finally, the RepGhost module is incorporated to significantly reduce the number of model parameters while maintaining high detection accuracy. Experimental results show that Pipe-YOLO achieves an mAP@0.5 of 91.42% and an mAP@0.5:0.95 of 69.11% on the dataset, with only 2.11M parameters. The results indicate that Pipe-YOLO improves detection accuracy while maintaining a low parameter count, making it suitable for lightweight pipeline defect detection. |
