Transfer Learning for Small Dataset Textile Applications
topic
Transfer learning enables effective deep learning model training for textile defect detection and classification with limited labelled data by initialising model weights from models pre-trained on large image datasets including ImageNet or industrial defect databases, then fine-tuning on 100 to 1000 textile-specific labelled examples, achieving classification accuracy that would require 10 to 100 times more training data with training from random initialisation.
Role
Overcomes the primary barrier to deep learning adoption in textile quality inspection where the cost and time of collecting and labelling thousands of defect images for each new fabric type and defect category makes full-dataset training impractical, with transfer learning enabling effective defect detection model deployment after labelling only hundreds of examples through reuse of visual feature recognition learned from large pre-training datasets.