Adaptive Recognition of Markings on Railway Wheels using Incremental Learning With YOLOv11

This paper investigates automated recognition of railway wheel marking symbols using YOLOv11 enhanced with incremental learning capabilities. The proposed approach aims to improve the classification of new symbols while mitigating catastrophic forgetting of previously learned markings. We employ a layer-freezing strategy (layers 4-23) to preserve core feature extraction capabilities while enabling adaptive finetuning of the classifier. Experimental evaluations demonstrate that stochastic gradient descent (SGD) optimization with a learning rate of 0.0005 achieves the most stable performance compared to alternative optimizers. Results indicate successful adaptation to new marking types while maintaining high accuracy (with only a $\sim 1 \%$ degradation) on original symbols, underscoring the effectiveness of the proposed incremental learning framework. The study highlights critical considerations for parameter selection in industrial computer vision applications, particularly in scenarios where marking methods evolve over time. The findings offer practical insights for deploying adaptive vision systems in manufacturing environments that require continuous model updates without full retraining.