Scholarlink

<p>The existing Image Coding for Machine (ICM) paradigm aims at simultaneously fulfilling both machine analytics and human perception needs by incorporating the performance constraint of downstream machine vision models. However, the intrinsic semantic gap among different vision tasks and the reliance on the performance of specific models, pose flexibility and generalization issues when handling unseen scenarios. To this challenge, this paper introduces a novel ICM paradigm that imposes an additional constraint on the reconstructed image from the Meta-Adversarial-Adaptation (MAA) perspective. To be specific, we first adopt an adversarial attack-for-good (AAFG) algorithm to generate subtle image perturbations that are almost perceptually lossless but significantly improve machine analytic performance. Thus, by adapting the target domain to the resultant image domain, the learned image codecs are customized for machine vision tasks, while still maintaining the image perception quality. Moreover, to minimize our method’s reliance on specific downstream models as much as possible for better generalization capacity, AAFG is performed in a meta-learning manner. A blanket of downstream models is employed to construct a <i>model set</i>. This set is subsequently fed into an iterative meta-adversarial augmentation process, aiming to capture the meta-information among these models and generate augmented images that generalize well. Extensive experimental results have demonstrated the effectiveness of our design in achieving satisfactory perceptual quality, improved machine analytics performance, and powerful generalization capacity regarding unseen downstream models, image domains, and object-centric tasks. Codes are available at <a href="https://github.com/YinKangSheng/ICM-MAA">https://github.com/YinKangSheng/ICM-MAA</a>.</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

Towards Generalized Image Coding for Machine Through Meta Adversarial Adaptation

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Xuelin Shen,
  • Kangsheng Yin,
  • Wenhan Yang

摘要

The existing Image Coding for Machine (ICM) paradigm aims at simultaneously fulfilling both machine analytics and human perception needs by incorporating the performance constraint of downstream machine vision models. However, the intrinsic semantic gap among different vision tasks and the reliance on the performance of specific models, pose flexibility and generalization issues when handling unseen scenarios. To this challenge, this paper introduces a novel ICM paradigm that imposes an additional constraint on the reconstructed image from the Meta-Adversarial-Adaptation (MAA) perspective. To be specific, we first adopt an adversarial attack-for-good (AAFG) algorithm to generate subtle image perturbations that are almost perceptually lossless but significantly improve machine analytic performance. Thus, by adapting the target domain to the resultant image domain, the learned image codecs are customized for machine vision tasks, while still maintaining the image perception quality. Moreover, to minimize our method’s reliance on specific downstream models as much as possible for better generalization capacity, AAFG is performed in a meta-learning manner. A blanket of downstream models is employed to construct a model set. This set is subsequently fed into an iterative meta-adversarial augmentation process, aiming to capture the meta-information among these models and generate augmented images that generalize well. Extensive experimental results have demonstrated the effectiveness of our design in achieving satisfactory perceptual quality, improved machine analytics performance, and powerful generalization capacity regarding unseen downstream models, image domains, and object-centric tasks. Codes are available at https://github.com/YinKangSheng/ICM-MAA.