License Plate Detection Utilizing Synthetic Data from Superimposition
Abstract: Machine learning projects are often constrained by data that is messy, scarce, sensitive or costly to produce. These are issues which could be mitigated by synthetic data. This thesis tries to improve Swedish license plate localization in images by synthesizing images through superimposition, a process that produces data cheaply and in abundance. A generative process algorithmically creates images of fake license plates which are glued on top of other images without any context awareness. The resulting images thus contain license plates where they normally don't exist, floating in the sky or across a person's face. To evaluate the usefulness of this method, a machine learning algorithm called YOLOv3 is trained on synthesized data and tested on real images containing Swedish license plates. Our findings show that training on synthesized data by itself is almost good enough to match the performance of training on publicly available real data containing international license plates. The best results are achieved by mixing real and synthetic data.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)