Optimized frameworks for deep learning on compressed images
(2025)
Files
Thesis__Karim_El_Khoury_DIAL.pdf
Open Access - Adobe PDF
- 2.98 MB
Details
- Authors
- Supervisors
- Macq, Benoît
- Abstract
- Global data consumption has surged 100-fold over the past 15 years, aligning with the rise of deep learning and driving a massive shift in big data analysis. This growth has significantly impacted fields such as medical imaging, video surveillance, and remote sensing. However, it has also challenged existing data management infrastructures, necessitating a balance between data constraints and analysis constraints. This balancing act translates to a trade-off between data compression efficiency and deep learning task performance. This thesis investigates strategies to better utilize data management infrastructures, aiming to minimize trade-offs between model performance and data compression. We explore this challenge through three distinct use cases: (1) improving deep learning segmentation robustness against image compression in medical imaging, (2) utilizing residual frames for object detection and tracking in video surveillance, and (3) developing a hybrid framework for bandwidth-constrained object detection in remote sensing. Our work reveals that fine-tuning deep learning segmentation and detection models on compressed data can be highly effective in both learning and filtering out compression artifacts. Additionally, our work demonstrates that models trained on compressed data can be easily integrated into existing frameworks for medical imaging, video surveillance, and remote sensing, reducing storage and bandwidth demands while maintaining high model accuracy.
- Affiliations
Citations
El Khoury, K. (2025). Optimized frameworks for deep learning on compressed images. https://hdl.handle.net/2078.5/240973