Image-based sequence recognition has been a long-standing research topic incomputer vision. In this paper, we investigate the problem of scene textrecognition, which is among the most important and challenging tasks inimage-based sequence recognition. A novel neural network architecture, whichintegrates feature extraction, sequence modeling and transcription into aunified framework, is proposed. Compared with previous systems for scene textrecognition, the proposed architecture possesses four distinctive properties:(1) It is end-to-end trainable, in contrast to most of the existing algorithmswhose components are separately trained and tuned. (2) It naturally handlessequences in arbitrary lengths, involving no character segmentation orhorizontal scale normalization. (3) It is not confined to any predefinedlexicon and achieves remarkable performances in both lexicon-free andlexicon-based scene text recognition tasks. (4) It generates an effective yetmuch smaller model, which is more practical for real-world applicationscenarios. The experiments on standard benchmarks, including the IIIT-5K,Street View Text and ICDAR datasets, demonstrate the superiority of theproposed algorithm over the prior arts. Moreover, the proposed algorithmperforms well in the task of image-based music score recognition, whichevidently verifies the generality of it.
展开▼
