Yolo- Smart Fabriscan for Fabric Defect Detection for Improving Missed Detections of Visually Similar Defects

Abstract

Fabric defect detection remains a significant challenge, necessitating innovative approaches to enhance detection efficiency and accuracy. The accuracy of fabric detection is better achieved by YOLO based models compared to other Deep Learning (DL) models. However, some categories remain challenging, such as holes and drops defects. Due to the similarities in colour and texture features of these defects, the model tends to misclassify these defects. In addition to that combined with external environmental factors such as changes in lighting conditions, can increase the visual similarity between these two defects, thereby impacting the model’s classification accuracy. To improve the typical YOLOv8 models for defect detection classification, customized SPP variations and attention is proposed in this paper to differentiate between fine-grained characteristics. Spatial Pyramid Pooling with Multi-scale Attention and Local Aggregation Network (SPP-MALAN) is proposed. Multi-scale Attention in SPP can improve automatically adjusts the weight distribution in the feature maps to focus more on key areas of the image. This is particularly effective in accurately identifying similar defects in various scales under complex environments like changes in lighting conditions. Local Aggregation Network improves the model’s representational capacity by efficiently integrating features from different layers of SPP. Convolution, Batch Normalization and SwiGLU (CBSG) replaces CBS to understand context and identify complex similarities between defects. It provides greater expressivity, leading to better accuracy and convergence of learning model. The proposed model is named as YOLO- Smart FabriScan. Experimental results demonstrate that the proposed YOLO-Smart FabriScan model achieves superior accuracy of 97.01%, 97.06%, and 98.1% on the TILDA 400, AITEX, and Fabric Stain datasets, outperforming all compared models.