基于少样本学习的图像修复技术综述

A Review of Image Inpainting and Restoration Based on Few-Shot Learning

摘要: 针对图像修复任务中普遍存在的数据获取难及退化模式复杂多变等挑战，少样本学习为模型在新场景下的快速泛化与适应提供了新思路。本文系统梳理了该领域的最新研究进展：首先探讨了元学习如何通过优化初始化策略实现对新任务的快速参数适配；其次分析了生成先验如何利用预训练模型的潜在空间，在极低样本下重构高频细节；同时阐述了自监督与对比学习如何挖掘数据内部相关性以增强表征稳健性，以及迁移学习如何将大规模域外知识有效映射至特定修复任务。尽管现有方法在纹理一致性与结构复原上取得显著突破，但仍受限于过拟合风险、未知退化建模及推理效率等瓶颈。展望未来，模型的高效微调及持续增量学习将是突破当前局限的关键方向。

Abstract: Research on few-shot learning-based image inpainting and restoration aims to tackle the challenges of completion and recovery under data scarcity and highly variable degradation distributions, thereby improving a model’s generalization and rapid adaptation to new scenarios. This review summarizes recent advances, covering meta-learning approaches that enable fast parameter adaptation with only a few samples; generative-prior methods that reconstruct fine details from extremely limited data; self-supervised and contrastive learning that enhance training stability via internal pseudo-supervision and representation constraints; and transfer learning that leverages pretrained knowledge and auxiliary-domain data for cross-domain adaptation. These methods have improved texture consistency, structural integrity, and perceptual realism, yet they still face challenges in controlling overfitting, modeling unseen degradations, selecting appropriate evaluation metrics, and achieving efficient inference. Finally, we highlight unified modeling of multiple degradations, efficient adaptation of foundation models, and continual/incremental learning as key future directions.