Fabric defect classification in textile manufacturing: enhancing CNN performance through hierarchical cross-stage attention using the ScienceDB dataset

Fabric defect classification has a significant impact on textile manufacturing efficiency and product quality. Traditional convolutional neural networks (CNNs) rely on static methods, limiting their ability to detect subtle fabric defects. This study proposes a novel Hierarchical Cross-Stage Attention (HC_CSA) model, which features dynamic multi-scale feature integration to enhance defect classification accuracy. Utilising the public ScienceDB dataset, which is applied for the first time in this research area, the proposed model was evaluated against established CNN architectures, including EfficientNet-B0, VGG19, AlexNet, and MobileNet. Results demonstrated superior performance, achieving 92.41% accuracy, 92.48% precision, 92.41% recall, and a 92.24% F1 score. These outcomes confirm that the proposed model effectively addresses the limitations of existing CNN approaches, providing enhanced adaptability and improved defect classification capabilities. These results were obtained under controlled patch-level evaluation conditions and provide a strong empirical foundation for the further development of automated textile quality monitoring systems.