Abstract: A three-stage strategy algorithm integrating SAM segmentation and YOLO technology was proposed to achieve high-precision surface defect detection through multi-model collaboration. A single-stage instance segmentation model was employed to detect regular and extremely small targets, while a single-stage object detector was used to identify extremely large targets. Combined with a large semantic segmentation model that provides mask support, these models collaborated through serial and parallel inference to accomplish the detection task. Results show that the proposed approach achieves an average detection accuracy of 97.3% in scenarios involving water-cooled wall images captured by wall-climbing robots. Key metrics exhibit outstanding performance, the average missed detection rate for defects is 0.75%, the false positive rate is 1.96%, the overall recognition accuracy is 97.3%, the missed detection rate is 0.75%, and the false positive rate is 2%. By integrating with unmanned aerial vehicle and wall-climbing robot technologies, this algorithm can effectively drive the development of boiler inspection systems toward automation and intelligence, providing a solution that balances both accuracy and efficiency for large-scale image defect detection of industrial equipment.