Handwriting Slant Determination Using Convolutional Neural Networks and Transfer Learning

Handwriting slant is one of the major characteristics utilized in forensic handwriting analysis. 1539 handwritten sample images were obtained from the online IAM Handwriting Database and labelled by a forensic expert for handwriting characteristics. Artificial Intelligence based Convolutional Neural Network classifiers were developed for classifying handwriting slant into three classes, viz. left, right and center. Transfer Learning approach from three pre-trained networks VGG16, ResNet50 and MobileNetV2 was attempted but it was found that Transfer Learning did not provide expected advantages, since presumably ImageNet database on which these networks were pre-trained does not have many images of handwritten samples. Therefore, robust and sufficient features could not be extracted using transfer learning approach. For handwriting images, it may be best to go for training from scratch until a pre-trained network which is trained on handwriting samples becomes available. A network trained from scratch using individual sentence images and L2-regularization gave the best performance. With 76% accuracy.