Speeding up document image classification

This work presents a solution by means of light Convolutional Neural Networks (CNNs) in the Document Classification task, essential problem in the digitalization process of institutions. We show in the RVL-CDIP dataset that we can achieve state-of-the-art results with a set of lighter models such as...

Descripción completa

Detalles Bibliográficos
Autor: Ferrando Monsonís, Javier|||0000-0002-2637-0961
Tipo de recurso: tesis de maestría
Fecha de publicación:2020
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/335901
Acceso en línea:https://hdl.handle.net/2117/335901
Access Level:acceso abierto
Palabra clave:Neural networks (Computer science)
Parallel programming (Computer science)
Clasificación de imágenes de documentos
Aprendizaje profundo
Sistemas paralelos
EfficientNet
BERT
Escalabilidad
TensorFlow
PyTorch
Document image classification
Deep learning
Parallel systems
Scalability
Xarxes neuronals (Informàtica)
Programació en paral·lel (Informàtica)
Àrees temàtiques de la UPC::Informàtica
Descripción
Sumario:This work presents a solution by means of light Convolutional Neural Networks (CNNs) in the Document Classification task, essential problem in the digitalization process of institutions. We show in the RVL-CDIP dataset that we can achieve state-of-the-art results with a set of lighter models such as the EfficientNets and present its transfer learning capabilities on a smaller in-domain dataset such as Tobacco3482. Moreover, we present an ensemble pipeline which is able to boost solely image input by combining image model predictions with the ones generated by BERT model on extracted text by OCR. We also show that the batch size can be effectively increased without hindering its accuracy so that the training process can be sped up by parallelizing throughout multiple GPUs, decreasing the computational time needed.