Recognition of Handwritten Swedish Sentences With Deep Learning

Abstract: This study attempts the task of handwritten text recognition within the context of the Swedish language. It examines the applicability of deep neural networks to comprehend handwritten Swedish texts, specifically leveraging the Labors Memory Dataset. The study employs Recurrent Convolutional Neural Networks (RCNNs) for encoding the visual features present inscanned handwriting and subsequently decoding them into textual representation. The model's output is checked against the human annotations of the data and evaluated using the character error rate measurement. Furthermore, this research explores the potential of transfer learning and ensemble methods to address the challenges posed by the limited volume of Swedish data and the high variance of deep neural networks. Across all methods explored, transfer learning from English OCR dataset combined with RCNN architecture achieved the best character errorrate of 0.166. Additionally, it was observed that ensembling with consensus sequence voting, while exhibiting promising results, comes with computational expenses that may not justify the accuracy trade off. Despite the efforts to address the shortcomings through image transformations, no improvements in performance were achieved. Further experiments were conducted to understand the limitations of the models and it was found out that they struggle notably when the data has a huge variance in scale.