Introduction and Background
Image compression is essential for reducing the size of visual data while preserving useful information. As image resolutions and sensor quality continue to increase, the gap between data volume and available storage or transmission bandwidth grows. Modern learned compression methods address this challenge by adapting to the statistical structure of images and often outperforming traditional codecs.
Satellite imaging amplifies these demands. Earth-observation satellites generate massive amounts of high-resolution, multispectral data, yet must operate under strict bandwidth and on-board compute limitations. General-purpose codecs are often insufficient, motivating compression methods tailored to the characteristics of remote-sensing imagery.
The COSMIC model is a recent neural compression approach designed specifically for satellite images, utilizing recent trends in computer vision and image compression such as diffusion models. This work aims to implement COSMIC and systematically evaluate different entropy models to better understand their impact on compression performance in the satellite domain.
Research Question
Main Research Question
How do different entropy models influence the performance of learned image compression?
Sub-questions
- Are certain entropy models generally performing "better" than others?
- Which entropy model is most suited for satellite images?
Tasks & Goals
- Literature Review
Study learned image compression and diffusion models - Data Preparation
Select benchmark satellite image data set - Model Development
Implement the paper "COSMIC: Compress Satellite Images Efficiently via Diffusion Compensation" - Training & Evaluation
Train and evaluate the model using different entropy models - Comparison & Analysis
Compare influence of entropy models on performance and efficiency - Thesis Documentation
Summarize findings, discuss limitations, and outline future research directions.
Requirements
- Strong Programming Skills:
Experience with Python, including libraries such as PyTorch and NumPy
- Deep Learning Knowledge:
Understanding of neural networks, especially CNNs and autoencoders. Completed prior coursework or projects using PyTorch or TensorFlow is essential. Experience with diffusion models is helpful but not mandatory
- Interest in computer vision and remote sensing methods
- Willingness to learn or deepen knowledge about diffusion models and learned image compression
References
Ziyuan Zhang, Han Qiu, Maosen Zhang, Jun Liu, Bin Chen, Tianwei Zhang, and Hewu Li. Cosmic: Compress satellite image efficiently via diffusion compensation. In Advances in Neural Information Processing Systems, pages 91951–91982. Curran Associates, Inc., 2024.