Prediction of Gender and Handedness from Offline Handwriting using Convolutional Neural Network with Canny Edge Detection

Handwriting classification based on a writer's demographics, such as gender and handedness, has been an essential discipline in forensic science and biometric security. Although there are already experts in forensic science called Forensics Document Examiners, their work can be affected due to a lack of efficiency and the risk of human errors. As there are only limited studies on handwriter demographics problems using Convolutional Neural Networks (CNN), this research implemented a system that predicts the gender, handedness, and combined gender-and-handedness of offline handwritten images from the IAM Handwriting iDatabase 3.0 using 2-Layer and 3-Layer CNN with Canny Edge Detection (CED). The researchers found that the base model 2L-CNN without CED had the best performance in the binary classes, gender, and handedness, with an overall accuracy of 68.5% and 89.75%, respectively. On the other hand, 3L-CNN without CED had the best average accuracy of 51.36% in the combined gender-and-handedness class. It was observed that Canny Edge Detection is not an effective preprocessing technique in handwriting classification as it worsened its counterpart’s performance, without CED, in most of the models.