Date of Award
2025
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
Committee Chair
Tathagata Mukherjee
Committee Member
Jacob Hauenstein
Committee Member
Vaidyanath Areyur Shanthakumar
Research Advisor
Tathagata Mukherjee
Subject(s)
Image compression, Remote-sensing images, Image processing--Digital techniques
Abstract
Neural compression methods claim near-lossless reconstruction capabilities based on traditional metrics such as PSNR and MS-SSIM. However, their impact on downstream machine learning tasks remains largely unexplored. This thesis evaluates three state-of-the-art CompressAI models (Cheng2020-anchor, bmshj2018-hyperprior and mbt2018 ) on high-resolution satellite imagery using semantic segmentation as a downstream task benchmark. We introduce the idea of algorithmic losslessness with respect to a downstream machine learning task and use it to benchmark existing state-of-the-art neural compression algorithms. We also compare the results with non-neural compression techniques. We conducted experiments on 11 urban satellite images using two segmentation approaches: MMSegmentation (7 classes) and FLAIR UNet (19 classes). Although all models achieved exceptional reconstruction metrics (0.985-0.996 MS-SSIM), semantic segmentation performance showed significant degradation. MMSegmentation IoU scores decreased by 12-44% between models, while FLAIR UNet demonstrated surprising robustness with only 0.3-6.8% performance loss. Our findings reveal a critical gap between perceptual reconstruction quality and functional task performance. Despite visually identical reconstructions, compression artifacts significantly affected segmentation accuracy, providing crucial insights for AI pipeline compression methods.
Recommended Citation
Simha, Abhishek Prasanna, "A study of neural compression techniques for image compression using a novel algorithmic lossless-ness criteria" (2025). Theses. 799.
https://louis.uah.edu/uah-theses/799