Improving CRNN with EfficientNet-like feature extractor and multi-head attention for text recognition

Text recognition is one of the most important and challenging tasks in image-based sequence recognition, which has various potential applications in real life. In this paper, we propose a novel convolutional-recurrent neural network (CRNN) for text recognition. Particularly, we adapt the EfficientNet architecture for extracting deep features and propose multi-head attention mechanisms to improve character localization. The experiments show that our EfficientNet-like feature extractor clearly outperforms other previous CNN feature extractors like VGG and ResNet. In overall, our proposed method yields competitive performance in comparison with other state-of-the-art approaches. Specifically, our F1-score is equivalent to top 3 on the ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction.