A comparison between Feed-forward and Convolutional Neural Networks for classification of invoice documents

Sammanfattning: Filing invoices under booking accounts can be a time-consuming task that could be alleviated by machine learning algorithms. There are two possible main methods for an algorithm to learn to classify such data: use a machine learning algorithm directly on the images, or extract words as tokens and use a machine learning algorithm on the set of words generated. This thesis tests these two methods with the use of two Artificial Neural Networks (ANN); a Feed-forward Neural Network(FNN) for a vectorized representation of the extracted word set and a Convolutional Neural Network(CNN) for the images. These models are also combined into ensembles made either entirely of FNN or CNN models, or equal parts of both. The findings in this thesis showed that a FNN trained on the word set turned out overall more accurate at classifying the invoices than a CNN trained on the images and that a mixed ensemble was more reliable when giving positive classification than the other ensembles.