A Novel Two-Way Transfer Learning Approach for Enhanced Fabric Fault Detection

Fabric defect detection is a crucial task in the textile industry, where early identification of faults ensures product quality and minimizes production losses. Conventional inspection methods, often manual and subjective, are insufficient for realtime and large-scale textile monitoring. While deep learning has significantly advanced defect detection performance, its effectiveness is hindered by the scarcity of labeled textile datasets and the domain gap between generic image features and fabric-specific textures. To overcome these limitations, this paper proposes a Two-Way Transfer Learning (TWTL)Approach for fabric fault detection, which leverages bidirectional knowledge transfer between a source domain (e.g., ImageNet) and a target domain (TILDA datasets). Our method integrates both forward and backward transfer mechanisms to enhance feature adaptability and defect classification accuracy. Experimental results on benchmark datasets such as TILDA demonstrate that the proposed approach outperforms traditional CNNs and one-way transfer learning models with notable improvements in performance metrics such as accuracy, precision etc. The optimized lightweight architecture of the model facilitates low-latency inference on edge devices, ensuring its applicability for real-time automated textile inspection in resource-constrained environments.