Multi Database Deep Learning Framework for Automated Textile Fabric Defect Detection and Benchmarking
The high-quality production of textiles, there is a need to detect any defects within fabrics accurately and efficiently. Traditional manual inspection techniques take time, are prone to errors, and are not applicable to mass production. In this paper, we present a multi-database deep learning approach that uses different databases of freely available and curated textile images to improve fabric defect detection. Our proposed approach uses a combination of different deep learning algorithms like CNN, ResNet, DenseNet, EfficientNet, and Vision Transformers to extract and classify features of defects on fabric samples. Another innovative aspect of our approach is the Hybrid Model incorporating multiple local models. Performance was assessed using Accuracy, Precision, Recall, and F1-Score measures on more than 10,000 labeled images of stained, hole-ridden, miswoven, and color-mismatched fabrics. It has been shown that the best results are achieved by the Hybrid Model with five local models, which yield higher Accuracy (91.0%), Precision (90.5%), Recall (90.2%), and F1-Score (90.4%) compared to other tested models. Cutting down the local models to three greatly affects their performance, which goes on to highlight the significance of having adequate models. For single models, DenseNet proved to be more efficient than others, while for the traditional ones, their consistency was relatively average. However, in terms of its process for inference, the framework supports large-scale implementation of the system in an industrial environment with real-time monitoring capabilities. Overall, this experiment clearly demonstrates the significant benefits of hybrid learning using multiple models and databases, and it can provide useful ideas for further research.