顶刊IJCV 2026！适配多重下游任务的通用图像增强框架

TL;DR - An IJCV-accepted paper presents a reinforcement-learning framework for adaptable image restoration across human-perception and machine-vision objectives. Its modular design handles mixed distortions and extends to new restoration tasks without retraining the entire system.

• A PPO-based controller selects and tunes operations for detected local and global distortions.
• A plug-in operator library supports denoising, texture restoration, color adjustment, and other enhancements.
• Task-specific evaluators reward perceptual quality or downstream detection and segmentation performance.
• Reported experiments cover perceptual restoration, object detection, segmentation, and few- or zero-shot motion deblurring.