A Comparison of Deep Learning CNN Architectures for Handwritten Joint Consonants

The majority of government buildings and workplaces in this age have transitioned to paperless processes, which boost efficiency, productivity, and—most importantly— environment friendliness. All paper documents must undergo optical character recognition (OCR) in order to be converted to digital or editable formats. Creating a Handwritten Gujarati Character Recognition model is more difficult when dealing with handwritten documents because various people have varied writing styles, sizes, and styles. Many handwritten character recognition systems are available for other languages, such as Hindi, Bengali and Devanagari, however there is still much work to be done for Gujarati language. The emphasis of the researcher is Gujarati Handwritten Joint Consonants, where information is gathered from 53 writers of various ages and genders for unique joint consonants consisting of 960 characters or classes. In this research, various CNN-based deep learning models for handwritten Gujarati joint consonant recognition are described. Additionally, talk about the accuracy rates obtained from various CNN model architectures, or more accurately, the pre-trained CNN models that are employed, such as vgg16, resnet50, inception_v3, efficientnetb2, and mobilenetv2. Finally, compare and discuss which model would be better based on the generated dataset.