A convolutional neural network model for robust classification of document-images under real-world hard conditions

Various studies have shown that convolutional neural networks (CNNs) can be successfully applied to classify documents types (by processing related document-images) from related document-images. Generally, document classes are categorized/differentiated through a similarity or not of their related respective structures. Although many scientific works have been published in this area of research, most of them do not reach the accuracy level required for most practical application scenarios (e.g., see digital office). This paper presents a new neural model based on convolutional neural networks for automatically and reliably detecting/classifying complex document types. A comprehensive benchmarking of our novel model with various other well-known CNN based classifiers clearly demonstrates that our model significantly outperforms all those other models by reaching an accuracy performance of 94.3%.