The impact of Data Augmentation on classification accuracy and training time in Handwritten Character Recognition

Sammanfattning: This bachelor thesis was conducted at the Royal Institute of Technology with the purpose of examining several combinations of data augmentation methods and their impact on classification accuracy of a CNN model. Further, the study investigates the time taken to train the model using the different data augmentation methods as to decide which ones have the best impact on classification accuracy in relation to their computational cost. The results indicate that the best classification accuracy is achieved when applying all provided augmentation methods although this leads to the longest training time. The study shows that a feasible alternative to the latter is to use a combination of augmentation methods which include Affine and Morphological transformations and Noise injection/removal since they yield a positive impact on classification accuracy while at the same time keeping the resource cost viable.