This paper presents a method of small object detection and applies it to warp yarn defect detection in warp knitting. Warp yarn defect is very challenging for existing small object detection methods. In the warping process, due to the fast speed and small size of warp yarns, the method requires high accuracy and real-time performance. In response to these challenges, we have made several changes to the current one-stage detection SOTA network--YOLOv8. An FPN attention module is designed to improve the effectiveness of spatial attention to small objects. A weighted pyramid is used to fuse feature maps at different scales effectively. The loss function is replaced with Inner-WIoU, which reduces geometric metric loss and focuses on general-quality samples to improve detection performance. The proposed method achieves a mAP of 99.4% for warp yarn defects. On the general dataset DIOR, the mAP reaches 88.5%. Meanwhile, the calculation amount of the proposed method is only 1/4 of YOLOv8m. However, in the DIOR dataset and the mixed dataset, the proposed method exceeds YOLOv8m in small objects. The proposed method effectively reduces missed detections and false detections in warp yarn defect detection tasks.