Deep learning based image compression

• Κύριος συγγραφέας: Ηλιοπούλου, Σοφία
• Άλλοι συγγραφείς: Iliopoulou, Sofia
• Γλώσσα: English
• Έκδοση: 2022
• Θέματα: Image compression; Deep learning; Autoencoders; Convolutional neural networks; Συμπίεση εικόνων; Βαθιά μάθηση; Αυτοκωδικοποιητές; Συνελικτικά νευρωνικά δίκτυα

Image compression is a research topic that has interested both the academic community and the business world for decades. The large size of media files combined with the popularity of social media and streaming services, render image compression necessary. However, it is imperative to maintain a decent image quality while reducing its size. Many methods have been developed and are still used, but in the last few years Artificial Intelligence has been making remarkable progress in the field. More and more image compression techniques choose to utilize Deep Learning now. This thesis first illustrates the range of application of Artificial Intelligence and then presents the related work in the field of Deep-Learning-based image compression. However, it is important to also study the traditional codecs like JPEG, since they are still used to this day, and they have become benchmarks for every compression algorithm. To this effect, the traditional image compression techniques are presented. Specifically, multiple methods are briefly explained, based on their effect on image quality. In addition, the different types of neural networks that are used for processing and compressing images, are presented. Their architecture and use are explained. Then, each part of the proposed implementation for image dimensionality reduction is shown. Finally, the performance of the developed system is evaluated and compared with other relevant attempts. In conclusion, the results show that the implementation of a Deep-Learning-based image compression method that is equal or even surpasses the traditional techniques in efficiency, is possible and high performance can be achieved. In a future work, the effect of different types of neural networks, as well as the improvement of the current method’s results will be studied.