Turkish Document Image Classification

Document image classification has gained extensive attention due to the rising number and types of scanned documents. Multi-modal architectures, processing image and text simultaneously, leverage the strengths of each modality. This study explores an efficient neural architecture for classifying scanned documents in a private company. The effectiveness of CNN-based deep learning and OCR algorithms in extracting textual and visual features is investigated. Different feature fusion methods are applied in the next stage to combine these extracted features. A multi-modal document image classifier is developed for companies managing a large number of scanned documents, delivering superior performance even with fewer and faint documents.