MSI-YOLO: an efficient model for multiscale fabric defect detection based on improved YOLOv11

Abstract

Fabric is the primary product of the textile industry, where defects are common and can directly affect quality and value of the final products. Timely detection of these defects is therefore essential for reducing losses and maintaining product quality. However, traditional defect detection methods often face limitations in efficiency and accuracy. To address these challenges, this study proposes the MSI-YOLO fabric defect detection algorithm based on You Only Look Once version 11 (YOLOv11). First, to enhance the model’s sensitivity to multiscale features and improve detection performance, the C3k2_SAConv module was developed by integrating Switchable Atrous Convolution (SAConv) concepts. Second, the Multiscale Dilated Attention mechanism is introduced to construct C2PSA_MSDA, strengthening the model’s ability to detect small defects while reducing the impact of complex backgrounds. Finally, Inner Intersection over Union (InnerIoU) was incorporated to replace Completed Intersection over Union (CIoU) as the regression loss function, further improving detection accuracy and reducing errors. The model’s performance was evaluated using precision, recall, mean average precision (mAP), and F1 Score. Experimental results demonstrate that, compared with the original YOLOv11n model, the improved MSI-YOLO model achieves increases of 3.7% in precision (P), 3.5% in recall (R), and 2.1% in mAP, with a frame per second (FPS) rate of 71.2. In addition, the model demonstrates excellent generalization ability, achieving an mAP at IoU threshold 0.50 (mAP@0.5) of 63.8% on the Alibaba Cloud Tianchi dataset, which is 6.4% higher than the baseline YOLOv11n model, and exhibits superior performance on other fabric defect datasets as well.