Evaluating the Performance of Different Text Detection and Recognition Models for Tyre Text

Computer Vision is a large part of artificial intel-ligence, specifically Optical Character Recognition. OCR helps by processing text and letters into digital content. One use case where Text recognition is not significant is tyre text. One challenge for the tyre text is the contrast between the text and its background. Using pre-trained text region detection and recognition models, this paper aims to recognize text using several models. Textsnake and CRAFT are used for text region detection, while ABINet, ASTER, and MORANv2 are used for text recognition. Results show Textsnake and ABINet to be the best. It was found that CRAFT had more correct detections by 5% on raw images and 1% on preprocessed images, but Textsnake was better for phrases and finding regions. As for Text recognition, it was observed that ABINet is better overall, with an accuracy of 40%, whilst ASTER is at 33%. Despite MORANv2 being the fastest by 25x, it was very inaccurate. In the end, using Textsnake and ABINet is the best out of all.