Analisis Akurasi pada Pengenalan Tanda Tangan Menggunakan Convolutional Neural Network dan Directed Acyclic Graph

Abstract

This research highlights challenges in authenticating signatures and aims to improve accuracy by comparing various parameters and CNN models. While previous studies have shown promising results using CNNs for signature recognition, the primary focus on training accuracy alone hasn't fully reflected optimal performance in real-world testing. Therefore, this study emphasizes the need to consider broader parameters and conduct more meticulous comparisons to achieve better outcomes. In efforts to enhance accuracy, the research introduces PCA and DAG methods to streamline the CNN training process. This approach aims to improve overall efficiency and accuracy. Through analyzing experimental results, researchers gained deeper insights into the performance of the developed models. Factors such as parameter adjustments and performance variations during training are discussed to strengthen understanding of the obtained results. Although training accuracy reached a satisfactory 97%, the testing results of only 73% indicate significant room for improvement. Thus, the analysis underscores the importance of refining parameters to enhance testing accuracy. By exploring broader parameters and employing structured DAG methods, this research offers potential for substantial performance improvements. Consequently, the study concludes that while achieving good training accuracy, substantial enhancements in testing accuracy can be attained through further research that expands understanding of optimal parameters and utilizes more structured methods in CNN training processes.