Efficient Document Classification Using Fused CNN-SVM Model

Storing massive amounts of data in physical form, issuing and printing documents, then filing them is impractical, inefficient, and antiquated, for those reasons, organizations have digitized document archiving by default. We propose a supervised learning approach that jointly integrates convolutional neural networks (CNN) with support vector machines (SVM) for real-time document classification in support of data extraction from electronically stored documents. This approach aims to improve classification accuracy when working with deformed or incomplete document images, for instance, containing signatures, marks, logos, or handwritten notes. In our research work, we extracted features using pre-trained architectures of Convolutional Neural Networks: EfficientNet, ResNet, VGG-16, and VGG-19. After collecting the features of fully linked layers from these architectures, we used the MRMR approach to select significant features. While this classification step has various features, ours used an SVM classifier instead of conventional CNN designs. It is equipped with the practical features that we produced from our CNNs. The results of the experiments showed that this strategy had far better classification ability than the general ways, with a beautiful efficiency rate of 95.39% in a shorter duration. Using an SVM classifier in combination with CNN-based efficient features answered firmly the problem of document classification.