An experimental investigation of MobileNet SSD on fabric defect detection

Abstract

Fabric defect detection plays a crucial role in ensuring product quality in textile manufacturing. This study presents an investigation into improving fabric defect detection using MobileNetV2 SSD as the baseline model and exploring various existing methods to further enhance performance. The project aims to improve detection accuracy without sacrificing speed and compares the effectiveness of focal loss, lightweight backbones, feature fusion methods, feature pyramid networks, and transfer learning. Extensive experimentation and evaluation were conducted using datasets from Roboflow on Google Colab and Kaggle. Experimental results showed that while focal loss improves accuracy, it also led to the problem of generating multiple bounding boxes for the same object and therefore not being used. Meanwhile, the concatenation module, SE ResNeXt50, and BiFPN were the top performers in their category. Transfer learning in smaller datasets yielded more significant improvements compared to larger datasets. MobileNetV2 BiFPN emerged as the best model that maintained fast speed while achieving high accuracy. The comprehensive insights garnered from this study can provide valuable guidance for future research in surface defect detection. Recommendations include building a custom dataset, conducting a deeper analysis of existing methods, and integrating a module for automatically detecting the optimal scale and aspect ratio.