This research project develops a deep learning-based model to recognize laser-etched serial numbers on copper surfaces in noisy industrial environments, addressing the challenge of a limited dataset. The experiments are split into two parts: background removal and character generation. Using Generative Adversarial Networks (GANs), background removal is performed unsupervised with public data, while character generation augments the dataset with stylized characters. The results show that fine-tuning with augmented data significantly outperforms background removal alone, leading to superior accuracy and reliability in serial number recognition.

Worarat Krathu

Pornchai Mongkolnam