Comparison of Deep Transfer Learning Against Contrastive Learning in Industrial Quality Applications for Heavily Unbalanced Data Scenarios When Data Augmentation Is Limited

Sensors, Journal Year: 2025, Volume and Issue: 25(10), P. 3048 - 3048

Published: May 12, 2025

AI-oriented quality inspection in manufacturing often faces highly imbalanced data, as defective products are rare, and there limited possibilities for data augmentation. This paper presents a systematic comparison between Deep Transfer Learning (DTL) Contrastive (CL) under such challenging conditions, addressing critical gap the industrial machine learning literature. We focus on galvanized steel coil classification task with acceptable vs. classes, where vast majority of samples (>95%) acceptable. implement DTL approach using strategically fine-tuned YOLOv8 models pre-trained large-scale datasets, CL Siamese network multi-reference design to learn robust similarity metrics one-shot classification. Experiments employ k-fold cross-validation held-out gold-standard test set images, statistical validation through bootstrap resampling. Results demonstrate that significantly outperforms CL, achieving higher overall accuracy (81.7% 61.6%), F1-score (79.2% 62.1%), precision (91.3% 61.0%) set. Computational analysis reveals requires 40% less training time 25% fewer parameters while maintaining superior generalization capabilities. provide concrete guidance when select over based dataset characteristics, demonstrating is particularly advantageous augmentation constrained by domain-specific spatial patterns. Additionally, we introduce novel adaptive framework integrates human-in-the-loop feedback domain adaptation techniques continuous model improvement production environments. Our comprehensive comparative offers empirically validated insights into performance trade-offs these approaches extreme class imbalance, providing valuable direction practitioners implementing systems limited, skewed datasets.

Language: Английский

Providing and evaluating a model for big data anonymization streams by using in-memory processing

Knowledge and Information Systems, Journal Year: 2025, Volume and Issue: unknown

Published: May 23, 2025

Language: Английский