Master Thesis: Learning Global-Local Representations for Efficient Satellite Image Compression

IABG Industrieanlagen-Betriebsgesellschaft mbH; München; Vor Ort; EUR 60.000 - 80.000; Vollzeit

Zusammenfassung

A leading company in the mobility and security sectors is offering a master's thesis opportunity focused on developing neural compression methods for satellite imagery. This project aims to enhance existing compression techniques by integrating domain knowledge, targeting efficient data handling for high-resolution satellite images. Candidates should possess strong programming skills, knowledge of deep learning, and be pursuing a master's degree in relevant fields.

Qualifikationen

• Strong knowledge of deep learning methods and related frameworks.
• Motivated Master student in relevant disciplines.
• Excellent communication skills in English.

Aufgaben

• Review state-of-the-art methods for neural-network-assisted image compression.
• Research and develop methods for embedding global context in local encodings.
• Benchmark proposed algorithm against compression algorithms.

Kenntnisse

Deep learning methods

Programming in Python

Communication skills in English

Ausbildung

Master student of Computer Science, Statistics, or related disciplines

Tools

PyTorch

Jobbeschreibung

Master Thesis: Learning Global-Local Representations for Efficient Satellite Image Compression

The IABG Innovation Centre is a development incubator for the IABG’s portfolio, which includes major trends in digitization, artificial intelligence, robotics, and sensor networks in the mobility and security sectors.
In recent years, Earth observation has become an essential component of various applications such as environmental monitoring, disaster response, and urban planning. The increasing resolution and frequency of satellite imaging has led to an exponential growth in data volume, making efficient image compression crucial for storage, transmission, and processing. While traditional compression techniques have been widely used, they often compromise on image quality or introduce artifacts. On the other hand, new deep learning-based approaches have shown great potential in achieving better trade-offs between compression ratio and perceived image quality.
However, these innovative approaches are mostly tailored to standard RGB images and lack explicit validation for satellite images, which exhibit different spatial and spectral characteristics. In the context of on-board compression to reduce downlink volume, efficiency becomes a core requirement. Given the sheer size of high-resolution satellite images, patch-based processing may be beneficial. Local-only methods risk losing critical global context, which is essential for preserving large-scale spatial patterns and semantic coherence. Furthermore, modeling global patterns across local patches miti-gates redundancy and enables more compact, low-entropy representations. This project aims to develop neural compression methods for satellite imagery, with a focus on global-local representations that can jointly capture broad contextual structure and fine local detail for optimal compression rates.

• Review state-of-the-art methods for neural-network-assisted image compression.
• Research and develop methods for embedding global context in local encodings.
• Evaluate the strengths and weaknesses of potential candidate Methods.
• Benchmark your proposed algorithm against compression algorithms.
• Motivated Master student of Computer Science, Statistics, or related disciplines
• Strong knowledge of deep learning methods and related frameworks (PyTorch preferred)
• Strong programming skills in Python
• Familiarity with satellite imagery is a plus
• Excellent communication skills in English

The goals of this master thesis are to develop a new method and novel perspective for compression of large image data, and to improve established compression methods by infusing domain knowledge. Furthermore, the thesis aims to estimate the maturity level for actual deployment of neural compression methods.