Character-based handwritten text transcription with attention networks

The paper approaches the task of handwritten text recognition (HTR) with attentional encoder–decoder networks trained on sequences characters, rather than words. We experiment lines from popular handwriting datasets and compare different activation functions for attention mechanism used aligning image pixels target characters. find that softmax focuses heavily individual while sigmoid multiple characters at each step decoding. When sequence alignment is one-to-one, able to learn a more precise decoding, whereas generated by much less precise. linear function obtain weights, model predicts character looking entire performs poorly because it lacks between source target. Future research may explore HTR in natural scene images, since capable transcribing without need producing segmentations or bounding boxes images.