Abstract:

Image inverse problems, which aims to reconstruct original images from degraded images or measurements observed by a given imaging system, is a hot topic in the recent development of signal processing. It has a wide range of application scenarios, such as compressed sensing, image denoising and image super-resolution etc. Recent years, the development of artificial intelligence has promoted the transformation of the field from analytical model-based methods to deep learning-based methods. However, whether in analytical model-based or deep learning-based methods, priors play a vital role in successfully solving image inverse problems. Different from existing related reviews which focus on introducing specific deep networks, this paper proposes a comprehensive and state-of-the-art review from a novel perspective of how to use priors in solving image inverse problems. After theoretical analysis and comparison of different methods, some future directions are finally put forward.

Key words: image inverse problem, analytical model, deep learning, priors

摘要：

图像逆问题是信号处理领域近年来非常热的一个研究课题，其目标在于从经由给定成像系统观测而来的退化图像或测量值中重构原始图像，具有广泛的应用场景，例如，压缩感知、图像去噪和图像超分辨率等。近年来，人工智能的发展推动了该领域从基于分析模型的方法到基于深度学习的方法的演变。然而，不论是在基于分析模型的传统方法还是基于深度学习的非传统方法中，先验信息的利用对于成功求解图像逆问题起着至关重要的作用。不同于现有相关综述研究侧重于介绍具体的深度网络，创新性地从如何利用先验信息求解图像逆问题的角度出发，归纳总结了该领域的研究现状，并对不同方法进行分析对比，最后展望了未来的研究方向。

关键词: 图像逆问题, 分析模型, 深度学习, 先验信息