Enhancement of convolutional neural networks algorithm for application form using GlobalMaxPooling in document verification system
Year: 2025
Type: article
Abstract: This study focuses on improving Convolutional Neural Networks (CNNs) to automate document verification system by utilizing CNN’s memory and computational complexity with adapting its structures to the specific characteristics of image data. Moreover, pooling is a crucial process for reducing the dimensionality of extracted features, a known key component of CNN Architecture. Thus, choosing the most appropriate pooling method is crucial across numerous computer vision architectures. Conventional pooling techniques like max pooling and average pooling have been extensively utilized for dimensionality reduction. However, both technique presents its own set of limitations such as loss of important details that are essential for tasks demanding high precision, while also diminishing the significance of key features by distributing attention across all values. This study presents the alternative pooling method of GlobalMaxPooling, aimed at capturing the most significant patterns and emphasizing critical patterns pertinent to document verification tasks. Using a dataset of 750 application forms, our results demonstrated an increase of significant improvement in detection accuracy, with the enhanced model achieving an accuracy of 77.2% compared to the existing model’s initial 20% accuracy. Furthermore, these findings emphasize the importance of effective pooling methods thereby strengthening the model’s capability for document verification, paving the way broader applications in automated systems requiring high precision and scalability.
DOI: 10.53378/isrr.166
Authors: Pamantasan ng Lungsod ng Maynila, Fabien Georgette Lapuz, Michael Carlos, Pamantasan ng Lungsod ng Maynila, Khatalyn Mata, Pamantasan ng Lungsod ng Maynila, Mark Anthony Mercado, Pamantasan ng Lungsod ng Maynila
Institution: University of the City of Manila
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410949442