Offline Author Identification using Non-Congruent Handwriting Data Based on Deep Convolutional Neural Network

Abstract

This investigation presents a novel technique for offline author identification using handwriting samples across diverse experimental conditions, addressing the intricacies of various writing styles and the imperative for organizations to authenticate authorship. Notably, the study leverages inconsistent data and develops a method independent of language constraints. Utilizing a comprehensive dataset adhering to American Society for Testing and Materials (ASTM) standards, a deep convolutional neural network (DCNN) model, enhanced with pre-trained networks, extracts features hierarchically from raw manuscript data. The inclusion of heterogeneous data underscores a significant advantage of this study, while the applicability of the proposed DCNN model to multiple languages further highlights its versatility. Experimental results demonstrate the efficacy of the proposed method in author identification. Specifically, the proposed model outperforms conventional approaches across four comprehensive datasets, exhibiting superior accuracy. Comparative analysis with engineering features and traditional methods such as Support Vector Machine (SVM) and Backpropagation Neural Network (BPNN) underscores the superiority of the proposed technique, yielding approximately a 13% increase in identification accuracy while reducing reliance on expert knowledge. The validation results, showcase the diminishing network error and increasing accuracy, with the proposed model achieving 99% accuracy after 200 iterations, surpassing the performance of the LeNet model. These findings underscore the robustness and utility of the proposed technique in diverse applications, positioning it as a valuable asset for handwriting recognition experts.