Abstract
Irrigation canal inspection requires accurate documentation of lining defects because cracks and perforated panels can increase seepage risk and reduce water-conveyance reliability. This study compared three YOLOv12 instance-segmentation variants, namely YOLOv12n, YOLOv12s, and YOLOv12m, for detecting two canal-damage classes: crack and perforated lining. The final dataset consisted of 1,200 real irrigation canal images collected from field canal sections under natural outdoor conditions. The annotation process produced 1,780 crack instances and 1,460 perforated-lining instances, resulting in 3,240 annotated damage objects. Images were assigned to training, validation, and test subsets using a class-proportion-preserving 80:10:10 allocation, with 960 images for training, 120 images for validation, and 120 images for testing. All models completed 100 epochs at 640 x 640 pixels using AdamW, a batch size of 16, and the same augmentation policy. Performance was evaluated using precision, recall, F1-score, mAP50, mAP50-95, inference time, FPS, matched-instance ROC-AUC, class-level performance, and object-level error summaries. YOLOv12m achieved the highest overall performance, with precision of 0.934, recall of 0.901, F1-score of 0.917, mAP50 of 0.948, and mAP50-95 of 0.721. YOLOv12n achieved the fastest inference speed on the RTX 3060 test workstation, reaching 128.2 FPS with an average inference time of 7.8 ms per image. These results establish an intra-generation accuracy-speed trade-off under a desktop-GPU environment; cross-generation and actual edge-device benchmarking remain outside the scope of this study.