Deep Learning-Based Optical Character Recognition for Identifying On-Label Printed Part Numbers of Used Automotive Parts: A Comparative Study of Open Source and Commercial Methods

This paper explores the use of deep learning-based optical character recognition (OCR) to identify part numbers for used automotive parts. It compares open source and advanced AI methods to commercial tools from Google, Amazon, and Microsoft. The study finds that fine-tuned open source models outperform commercial services, especially for complex part numbers unrelated to any language structure. The preferred open source method, MaskedTextSpotter, is fine-tuned with image data from old vehicle and electrical parts, captured by a smartphone and 2D barcode scanner. Additionally, a new data augmentation method, CharChan, is introduced, replacing detected characters with random examples for better character recognition. The experiments demonstrate the efficacy of deep learning-based OCR for automotive part number identification.