LoRA 與其變體應用於深度影像壓縮模型之效益探討

Abstract

隨著深度學習技術的發展，深度影像壓縮模型已大量應用於日常生活中。然而， 傳統完整微調 (Full Fine-Tuning, FT)方式需耗費大量時間與計算資源，限制了可變率影像壓縮 (Variable-Rate Image Compression) 技術在資源受限設備上的應用。為此，本研究以系統性方式探討參數高效微調技術 LoRA (Low-Rank Adaptation) 及其變形體 (AsyLoRA 與 DoRA) 在可變率影像壓縮中的應用效能。 實驗結果顯示，LoRA 及其變形體相較於傳統完整微調方式，能將需更新參數量降至原先需求約 1% 以下，並將記憶體儲存需求從原先的 877 MB 降至僅約 4 MB，同時還能維持良好的壓縮功效。特別是 DoRA，在僅略增儲存開銷前提下，提供比 LoRA 更佳且更穩定的效能; 而 AsyLoRA 雖然相較於DoRA效能稍低，但所需記憶體量最低，特別適合應用於極低資源的環境。 總結而言，這些低秩調適方法在可變率影像壓縮中展現出高度的應用潛力，本研究為未來在資源受限情境下可變率影像壓縮的應用奠定了初步的基礎

With the development of deep learning technologies, deep image compression models have been widely applied in daily life. However, traditional full fine-tuning (FT) requires significant time and computational resources, limiting the practical deployment of variable-rate image compression techniques on resource-constrained devices. To address this, this study systematically investigates the effectiveness of parameter-efficient fine-tuning techniques, namely LoRA (Low-Rank Adaptation) and its variants (AsyLoRA and DoRA), in variable-rate image compression. Experimental results show that LoRA and its variants can reduce the number of trainable parameters to about 1% of that required by traditional FT and lower storage requirements from 877 MB to approximately 4 MB, while still maintaining good compression performance. In particular, DoRA offers better and more stable performance than LoRA with only a slight increase in storage overhead. Although AsyLoRA shows slightly lower performance compared to DoRA, it achieves the lowest memory consumption, making it especially suitable for extremely resource-limited scenarios. In summary, these low-rank adaptation methods exhibit great potential for variable-rate image compression and lay a solid foundation for future applications on resource-constrained devices.