Gender prediction from handwritten digits: Comparison and analysis of deep learning approaches

For a long time, the digitization and storage of information in various fields has become increasingly important. In particular, there is a significant demand for the conversion of printed, handwritten or image data into an editable format. Various issues arise both in the digitization process and in the recognition of such information. For researchers, this area is very attractive. There are various technological approaches to extract and store textual data from image files. Optical Character Recognizer (OCR), a widely recognized technology, converts text from scanned documents or images into digital text that can be edited and used for storage. However, character features, lack of whitespace, axis shifts in text, and especially handwritten ones, pose various problems and challenges as a result of the image quality in scanned documents. The use of artificial intelligence-based techniques has improved the accuracy of character recognition. Among these techniques, Convolutional Neural Networks (CNNs) are preferred for computer vision problems. In this study, handwritten digits were collected from students in the Turkish National Education System and the data were appropriately labeled. The dataset aims to identify students' handwriting and gender. Five different well-known CNN models (Customized CNN, Resnet-50, InceptionV3, Vgg16, Vgg19) were used. The results obtained from these models are compared.