@misc{Filinger2020, type = {Master Thesis}, author = {Filinger, Andreas}, title = {Improving Automatic Meter Reading using Data generated from Unpaired Image-to-Image Translation}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:860-opus4-2413}, school = {Hochschule f{\"u}r Angewandte Wissenschaften Landshut}, pages = {31}, year = {2020}, abstract = {The automatic reading of water readers or electric readers is a challenging problem that may be solved using deep neural networks (DNNs). Unfortunately, the training of DNNs requires large sets of labeled data that are expensive to obtain. We alleviate this problem by generating artificial training data using unpaired image-to-image translation based on CycleGANs, ie a special form of Generative Adversarial Networks. We have implemented and evaluated a pipeline for generating artificial meter images using a medium sized dataset (6000 images) of water meter images. Our results confirm that this methodology is indeed effective, especially on small initial datasets.}, subject = {Generative Adversarial Networks}, language = {en} }