Worddeepnet: handwritten gurumukhi word recognition using convolutional neural network

Deep learning models are considered a revolutionary learning paradigm in artificial intelligence and machine learning, piquing the interest of image recognition and computer vision experts. Because deep learning models have gained popularity and improved outcomes in the literature, this work provides a deep learning method based on a holistic approach to recognize offline handwritten Gurumukhi words. The holistic approach to word recognition treats a word as a separate entity rather than its component letters. Three characteristics are extracted from word pictures to train a Convolutional Neural Network (CNN), namely, zoning, diagonal, and centroid. Five performance measures are used to assess trained CNN performance, namely, Accuracy, True Positive Rate (TPR), False Positive Rate (FPR), Root Mean Square Error (RMSE), and Area Under Curve (AUC). The proposed model is trained and assessed using a 40,000 words benchmark dataset based on 70:30 partitioning technique, in which 70% of the data is used to train the model and 30% of the data is used to test the trained model. To assess the efficacy of the suggested technique, a fivefold cross validation process is performed. Using the partitioning method and cross-validation approach, the best accuracy rates of 95.11% and 94.96% are obtained after 30 epochs, respectively which surpassed the existing state-of-the-art offline handwritten Gurumukhi word recognition systems.