MULTI-SCALE FEATURE FUSION VIA VIT AND EFFICIENTNET: A DUAL-BRANCH APPROACH TO INDUSTRIAL FABRIC INSPECTION.

Industrial fabric inspection for richly patterned textiles of complex textile such as Ankara fabrics remains a challenging task due to the coexistence of fine-grained defects and complex global motifs. Conventional vision systems and single-architecture deep learning models often fail to simultaneously capture microscopic anomalies and long-range structural dependencies, resulting in missed defects and high false alarm rates. This study presents a multi-scale feature fusion framework based on a dual-branch architecture that integrates EfficientNet-B0 and Vision Transformer (ViT-Base16) for robust industrial fabric inspection. The EfficientNet branch is responsible for extracting high-resolution local texture features critical for detecting subtle defects such as stains, holes, and yarn inconsistencies, while the Vision Transformer branch captures global contextual patterns and long-range dependencies inherent in complex fabric designs. A fusion mechanism is introduced to combine these complementary representations, enabling the model to learn both local and global feature hierarchies effectively. The proposed system is trained using a one-class classification paradigm on 2,000 defect-free samples, allowing it to model normal fabric distributions and identify anomalies without requiring extensive defective datasets. Experimental evaluation demonstrates that the hybrid model significantly outperforms baseline architectures, including ResNet-50 and standalone variants, achieving a ROC-AUC of 97.4%, Recall of 96.8%, Precision of 97.1%, and an F1-Score of 96.9%, alongside a low false positive rate of 0.9%. These results confirm the effectiveness of multi-scale feature fusion in enhancing detection accuracy and reliability. Furthermore, the model maintains competitive inference speed suitable for real-time deployment in industrial environments. The findings highlight that combining convolutional inductive biases with transformer-based global reasoning provides a powerful and scalable solution for automated textile quality control. This work advances the state of the art in fabric defect detection and offers practical implications for intelligent manufacturing systems