Date of Award
2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
Committee Chair
Tathagata Mukherjee
Committee Member
Letha Etzkorn
Committee Member
Jacob Hauenstein
Research Advisor
Tathagata Mukherjee
Subject(s)
Image compression, Remote sensing--Data processing, Image processing--Digital techniques
Abstract
Although advanced neural image compression optimized for human perception- oriented metrics such as PSNR and MS-SSIM demonstrates promising rate-distortion performance, the effectiveness of such approaches for downstream tasks, e.g., semantic segmentation, remains uncertain. Our large-scale empirical study of remote sensing imagery shows the quantization bottleneck in neural codecs destroys high-frequency textures relevant for downstream segmentation at a disproportionate rate, resulting in large and unpredictable drops in mean Intersection over Union that pixel-fidelity metrics cannot predict. In this thesis, we tackle this problem by implementing Weight- Decomposed Low-Rank Adaptation (DoRA) only on the synthesis transform, without re-training the codec or modifying the bitstream. It learns the low-rank directional updates by label-free distillation from a static FLAIR-INC segmentation teacher via boundary and saliency weighting and multi-depth feature matching. The modified MBT2018 q8 decoder improves the consistency mIoU from 0.643 to 0.723 (+8.04 points), while the PSNR and MS-SSIM remain unchanged. It outperforms all JPEG, JPEG-2000 and WebP operating points below 12 bpp.
Recommended Citation
Gautam, Ashish, "Mitigating reconstruction loss in neural compression for downstream tasks using weight-decomposed low-rank adaptation" (2026). Theses. 844.
https://louis.uah.edu/uah-theses/844