VGG16: Offline handwritten devanagari word recognition using transfer learning
Garg, Naresh Kumar
DOI: https://doi.org/10.1007/s11042-024-18394-7
IF: 2.577
2024-02-11
Multimedia Tools and Applications
Abstract:Word recognition has garnered significant attention due to its wide-ranging real-time applications, including postal automation, signature verification, bank check processing and writer identification. Handwritten word recognition poses a formidable challenge in computer vision, primarily due to the diverse writing styles exhibited by individuals. Presently, researchers are increasingly focused on developing holistic word recognition approaches. This holistic approach treats entire words as single, indivisible entities for the purpose of feature extraction and recognition, rendering it well-suited for identifying words with overlapping characters. In this study, authors have proposed a holistic approach for recognizing offline handwritten words in the Devanagari script, employing VGG16 (Visual Geometry Group) as a versatile feature extractor. Three classifiers, namely Gaussian Naive Bayes (Gaussian NB), eXtreme Gradient Boosting (XGBoost) and Random Forest (RF), are considered for classifying these word images. Based on the extracted features and the aforementioned classifiers, words are categorized into one of 50 distinct word classes. Various performance evaluation metrics, including Accuracy, Precision, Recall, F1-score and Area Under the Curve (AUC), are employed to assess the effectiveness of the proposed system. It is worth noting that the Random Forest (RF) based classification technique achieved exceptional accuracy, with impressive recall, high precision and a remarkable F1-score. Additionally, the XGBoost based classification exhibited comparable performance, achieving the highest possible AUC in this study. Furthermore, to underscore the significance of this work, authors have presented a comparative analysis of the proposed approach with existing state-of-the-art methodologies.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering