Few-Shot Domain Adaptation for Learned Image Compression
This addresses the generalization issue in learned image compression for applications requiring adaptation to new domains with minimal data, though it is incremental as it builds on existing LIC schemes.
The paper tackles the problem of learned image compression models performing poorly on out-of-training-domain images by proposing a few-shot domain adaptation method using plug-and-play adapters, achieving performance comparable to H.266/VVC intra coding with only 25 target-domain samples and matching full-model finetune with less than 2% of parameters transmitted.
Learned image compression (LIC) has achieved state-of-the-art rate-distortion performance, deemed promising for next-generation image compression techniques. However, pre-trained LIC models usually suffer from significant performance degradation when applied to out-of-training-domain images, implying their poor generalization capabilities. To tackle this problem, we propose a few-shot domain adaptation method for LIC by integrating plug-and-play adapters into pre-trained models. Drawing inspiration from the analogy between latent channels and frequency components, we examine domain gaps in LIC and observe that out-of-training-domain images disrupt pre-trained channel-wise decomposition. Consequently, we introduce a method for channel-wise re-allocation using convolution-based adapters and low-rank adapters, which are lightweight and compatible to mainstream LIC schemes. Extensive experiments across multiple domains and multiple representative LIC schemes demonstrate that our method significantly enhances pre-trained models, achieving comparable performance to H.266/VVC intra coding with merely 25 target-domain samples. Additionally, our method matches the performance of full-model finetune while transmitting fewer than $2\%$ of the parameters.