Deep Network with Pixel-Level Rectification and Robust Training for Handwriting Recognition

CNN-BiLSTM model for English Handwriting Recognition: Comprehensive Evaluation on the IAM Dataset

“…Our model is able to compete with the similar approaches proposed in [4,47]. However, we were not able to fully duplicate their results due to the differences including (1) extra data they used in training, (2) the language model they employed to support the deep learning network output, and (3) the lexicon their methods select decoded words from.…”

“…Affine transformations such as rotating, scaling, and shearing are heavily applied and are shown to be effective methods for mimicking handwriting styles [9,36,47]. More complex augmentation techniques are also proposed.…”

“…Most of the studies have generated their own synthetic data, either in word or line level, and experimentally shown their effectiveness. While some have synthesized a few millions of handwriting images using words or sentences from known, large corpora [9,22,47] synthetic data generated for scene-text recognition problem [24], or cropped lines from pdf files containing handwriting text [27]. However, though the above methods report performance increase, none of them have published their dataset, which leaves the problem unsolved for other researchers.…”

“…Although data augmentation for visual tasks has been explored extensively [40], there are a couple of studies that proposed novel approaches to augment handwriting images [28,41,46], while others utilized common methods such as affine transformations [9,36,46,47].…”

“…Following [9,36,47], we employed affine conversions including shear and rotation to generate natural looking handwriting images. Moreover, we applied elastic distortion [46] and geometric transformations [28] to further increase the handwriting variability.…”