Fabric defect detection serves as a critical stage in textile quality control processes. To substantially improve detection accuracy and enhance real-time performance, this paper proposes a cloud-edge computing-based fabric defect detection system. The system implements an optimized PGHGNetV2 backbone network on edge devices, integrated with the SPPELAN module and Slim-neck structure, effectively reducing computational redundancy and model parameters while improving small defect detection capability. During the final processing stage, the system incorporates the iRMB attention mechanism to further enhance small defect feature extraction and suppress background noise interference. Experimental results confirm that the proposed method achieves an optimal balance between accuracy and efficiency. Compared to mainstream detection models, our approach maintains a high mAP@50 of 98.6% while reducing computational complexity by 23.6%. The detection performance reaches 5.86 FPS on embedded NVIDIA Jetson Nano platforms, conclusively validating the effectiveness of the proposed network architecture improvements and confirming their practical implementation value in industrial applications.