CAFPN: A Network for Offline Text Line Recognition with Attention-based Multi-stage Feature Fusion and Context Parser

Present approaches to handwritten recognition primarily focus on enhancing the capabilities of feature extractors and context parsers to resolve issues. However, the diversity and ambiguity in handwritten fonts often mislead Convolutional Neural Networks (CNNs) during the image feature extraction process. Increasing the depth or altering the architecture of feature-extracting CNNs fails to effectively address this problem. Furthermore, context parsers based on Long Short-Term Memory (LSTM) face inefficiencies due to their sequential nature, which slows down the overall inference speed of the model. In this paper, we propose a novel method based on self-attention for multi-stage feature fusion, which introduces two types of global features. These are then automatically weighted and chosen by the model's attention mechanism to assist the final layer, effectively mitigating the feature extractor's susceptibility to misguidance. Additionally, we introduce a self-attention-based context parser that operates in parallel and delivers superior long-distance parsing results, addressing the slow inference speed issue inherent in LSTM-based context parsers. Extensive experimental results demonstrate consistent performance improvements on five public datasets.