JPEG2000 image compression for deep learning of convolutional neural networks

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Introduction: The Deep Learning of Convolutional Neural Networks has become quite a popular and really effective approach to tackle complex problems especially with images. Also, with the recent improvements to the performance of GPUs the ease of training DL models has increased and their adoption is now broader than ever. However, a big issue still remains; those models require a lot of data to train on and the sizes of the datasets are increasing at the same rate as the complexity of those models. This is why compression has become a necessity in order to reduce the storage and bandwidth requirements. Another issue is that standard codecs were thought and optimised for the human visual system and could be sub-optimal when used for a computer-vision task. In order to address this issue, this master thesis will explore some of the parameters of a specific compression codec and will aim to bring light to their effect on the performance of a computer-vision task. Material and methods: In order to see if some behaviours are generalisable, two bundles each containing a dataset and a model were used. The first bundle used YOLOv4 on a face-mask detection dataset to have a first glance at the influence of the compression as well as to try some ideas. The second bundle used Tiny-YOLOv4 with the UA-DETRAC vehicle dataset in order to get fast training times and this way be able to explore as many parameters of JPEG2000 as possible. Results: The first set of experiments highlights that it could be possible to compress a dataset heterogeneously in order to maximise the overall compression ratio while retaining good performance. The main exploration of the JPEG2000 parameters done in the second set of experiments yields mixed results. Optimising the parameters is difficult because of the random nature of the results but is possible, even though overall this set of experiment did not bring major improvements to the performance. Conclusion: This field is still rather unexplored and future works could have a tangible impact on how the datasets of tomorrow are stored and used.