Scholarlink

<p>This paper proposes a multi-task framework for learning-based image compression in which multiple tasks share a common latent representation while preserving compatibility with a single frozen reconstruction decoder. Unlike existing approaches that retrain both encoder and decoder for each task configuration, the proposed method adapts only the encoder and task-specific heads, maintaining decoder standardization and interoperability. Built upon the HiFiC codec, the framework supports additional tasks such as image super-resolution and facial feature extraction from the compressed domain. An adaptive multi-task loss balances compression efficiency and task performance. Experiments at different bitrates demonstrate that heterogeneous tasks can be integrated within a shared latent space while preserving competitive rate-distortion performance. These results support the development of interoperable AI-based compression systems for both visual reconstruction and downstream inference, under a fixed, shared decoder.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Beyond Compression: Revisiting the Encoder for Multi-Task Learning in AI-Based Image Compression

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Mohamed El-Amine Bellebna,
  • Sid-Ahmed Berrani,
  • Mohamed Riadh Temmar,
  • Jean-Luc Dugelay

摘要

This paper proposes a multi-task framework for learning-based image compression in which multiple tasks share a common latent representation while preserving compatibility with a single frozen reconstruction decoder. Unlike existing approaches that retrain both encoder and decoder for each task configuration, the proposed method adapts only the encoder and task-specific heads, maintaining decoder standardization and interoperability. Built upon the HiFiC codec, the framework supports additional tasks such as image super-resolution and facial feature extraction from the compressed domain. An adaptive multi-task loss balances compression efficiency and task performance. Experiments at different bitrates demonstrate that heterogeneous tasks can be integrated within a shared latent space while preserving competitive rate-distortion performance. These results support the development of interoperable AI-based compression systems for both visual reconstruction and downstream inference, under a fixed, shared decoder.