One-stage Vehicle Engine Number Recognition System

Abstract

This study proposes a one-stage vehicle engine number recognition system which avoids using the traditional three-stage recognition procedures of positioning, segmentation, and then character recognition, without the needs of image preprocessing procedures, we directly locate and recognizes the text targets in the vehicle engine image. The experiment using 926 labeled images via transfer learning to train our prediction model and then using this prediction model to test another 2310 unlabeled images, the overall accuracy achieved 99.48% and the execution time for recognize a single image is 234ms.