FA-YOLO: A High-Precision and Efficient Method for Fabric Defect Detection in Textile Industry

The automatic defect detection for fabric images is an essential mission in textile industry. However, there are some inherent difficulties the of images, such as complexity background and highly uneven scales defects. Moreover, trade-off between accuracy speed should be considered real applications. To address these problems, we propose a novel model based on YOLOv4 to detect defects called Feature Augmentation YOLO (FA-YOLO). In terms network structure, FA-YOLO adds additional head improve ability small builds powerful Neck structure enhance feature fusion. First, reduce information loss during fusion, perform residual augmentation (RFA) features after dimensionality reduction by using 1×1 convolution. Afterward, attention module (SimAM) embedded into locations with rich adaptation complex backgrounds. Adaptive spatial fusion (ASFF) also applied output filter inconsistencies across layers. Finally, cross-stage partial (CSP) introduced optimization. Experimental results three industrial datasets, including Tianchi dataset (72.5% mAP), ZJU-Leaper (0.714 average F1-score) NEU-DET steel (77.2% demonstrate proposed achieves competitive compared other state-of-the-art (SoTA) methods.