Attention-Driven Handwritten Identification: A Segmenation-Free Framework with Diffusion Based Data Augmentation

Human handwritten character identification is a highly active research field in computer vision. Handwritten recognition is a challenging task due to the complex nature of writing. As a result, many applications still depend on the manual transcription of handwritten documents into digital formats. Deep learning-based models have achieved remarkable performance in image recognition tasks, but they need large amounts of data, which provides less accuracy for languages with minimum datasets. Therefore, finding solutions to overcome this issue is essential for improving the accuracy of handwritten text recognition across different languages.To overcome the difficulties and limitations of deep learning-based models, a novel approach is employed: diffusion probability-based data augmentation and attention-based sequence-to-sequence models are used to identify the handwritten text. Data augmentation is a technique employed to increase the amount and diversity of available data. Attention mechanisms help the model to deal with long and complex sequences and improve the accuracy and efficiency of the model.Experimental results on the IAM dataset reveal that the LSTM-CNN-based model achieved a character-level accuracy of 94.00% while utilizing only half of the dataset (51.83%) for training. Notably, this performance is comparable to state-of-the-art models, demonstrating the model’s efficiency and effectiveness even with a reduced volume of training data. The diffusion-based model, on the other hand, generated realistic and diverse instances of handwritten text with low resource consumption. Therefore, the diffusion-based data augmentation and attention-based LSTM model demonstrated a more effective and efficient approach for handwriting recognition.