Deep learning for classification of scanned documents : What is important?

Sammanfattning

This thesis investigates how the quality and quantity of data affect the performance of deep learning models in multimodal document classification. It explores when the input documents are represented both as images and as OCR-extracted text from the image. While deep learning models have shown strong performance on clean, large-scale datasets, many applications often involve noisy data and limited annotations, especially in sensitive domains such as healthcare. To study these effects, a series of controlled experiments are designed that simulate degraded data conditions. The training dataset size is reduced and artificial corruptions are applied to both visual image resolution and textual OCR quality modalities. The impact of these manipulations are evaluated both individually and in combination across several degradation levels, using the Ryerson Vision Lab Complex Document Information Processing (RVL-CDIP) dataset as a proxy for medical document collections. The results indicate that image degradation has a slightly greater impact on model performance than reduced text quality or training set size alone. When both modalities are degraded, performance drops significantly, even when using a very large dataset size. These findings suggest that deep multimodal models are particularly sensitive to visual input quality. But as long as at least one modality remains informative, the model can maintain reasonable accuracy even with limited data. Implementation details, evaluation procedures, and a reproducible pipeline are provided to support further research in low-resource settings.