AI-Based Real-Time Fabric Defect Detection Using an Enhanced SSD with Bidirectional Feature Pyramid Networks

Fabric defect detection is one of the most significant challenges in the textile industry due to its critical role in quality assessment and management. Conventional single-stage detectors often capture small-scale and low-contrast defects inadequately. On the other hand, high-accuracy two-stage methods suffer from excessive computational complexity. This study proposes an improved Single Shot MultiBox Detector (SSD) model by replacing the feature map layer with the Bidirectional Feature Pyramid Network (BiFPN) from EfficientDet. Also, Bayesian optimization is utilized to systematically tune the related hyperparameters, which improves convergence stability and detection performance without manual intervention. Performance evaluation involves a trade-off between mean average precision at IoU 0.5 (mAP50) and execution time or frames per second (FPS), given that fabric defect detection requires the rotation of fabric motors or rollers. Experiments on a fabric defect dataset demonstrate that the proposed SSD-BiFPN framework outperforms baseline SSD models when it comes to precision, recall, and mean average precision, particularly for small and irregular defects. Additionally, the proposed architecture demonstrates satisfactory real-time performance when implemented on an NVIDIA Jetson Nano platform, highlighting its appropriateness for edge-based industrial inspection scenarios.