Handwritten Character Recognition using Dynamically Mapped CNN Approach
N R Sreekumar, P Rakshitha, S Harshitha, N. Shobha Rani · 2023
This paper introduces a novel approach to developing a handwriting recognition system by leveraging deep learning techniques. The proposed method involves several stages. Firstly, a dataset is obtained from Kaggle, followed by a thorough data-cleaning process to eliminate unreadable samples. The data is then preprocessed by converting the images into the grayscale format, resizing them to a width of 256 and a height of 64, and finally normalizing the pixel values to fit within the range of [0, 1]. Subsequently, the dataset has been split into training sets and testing sets. with appropriate labels prepared for the CTC loss function. The model architecture is constructed by combining convolutional neural networks (CNNs) for feature extraction with sequence layers containing bidirectional LSTM (bi-LSTM) units. The model is trained using the training set and then evaluated using the testing set, utilizing n-gram analysis as the evaluation metric. The evaluation results show that the developed handwriting recognition system is efficient and successful.