arrow
返回

Reverse Image Filtering Using Total Derivative Approximation and Accelerated Gradient Descent

delete2022-01-01
delete3
delete
OA
AI
F
Fernando Galetto *
G
Guang Deng
DOI:10.1109/ACCESS.2022.3225411delete
delete原文链接
delete原文求助
delete分享
delete收藏
摘要

摘要

En 中文
In this paper, a new problem of reverse image filtering is addressed. The problem is to reverse the effect of an image filter given the observation b = g(x). The filter g is modelled as an available black box. An inverse method is proposed to recover the input image x. The key idea is to formulate this inverse problem as minimizing a local patch-based cost function. To use gradient descent for the minimization, total derivative is used to approximate the unknown gradient. This paper presents a study of factors affecting the convergence and quality of the output in the Fourier domain when the filter is linear. It discusses the convergence property for nonlinear filters by using contraction mapping as a tool. It also presents applications of the accelerated gradient descent algorithms to three gradient-free reverse filters, including the one proposed in this paper. Results from extensive experiments are used to evaluate the complexity and effectiveness of the proposed algorithm. The proposed algorithm outperforms the state-of-the-art in two aspects. (1) It is at the same level of complexity as that of the fastest reverse filter, but it can reverse a larger number of filters. (2) It can reverse the same list of filters as that of a very complex reverse filter, but its complexity is much lower.
Keyword:
Maximum likelihood detection
Inverse problems
Nonlinear filters
Filtering algorithms
Minimization
Cost function
Image filtering
Inverse filtering
optimization
accelerated gradient descent
total derivative

期刊

IEEE Access 封面图
IEEE Access
IF:
3.6
论文数:
9.8W
被引数:
29.4W

机构

L
La Trobe University
学者数:
1.1W
论文数: 1.1W
被引数: 1.5W
引用论文

引用论文

err分享
err收藏
Iridium-catalyzed crossed aldol coupling and a new concept to form an Ir–Si species
err2000-02-01
err0
PREAI
errIsamu Matsuda; Yuki Hasegawa; Tatsuya Makino; Kenji Itoh
err分享
err收藏
err分享
err收藏
err分享
err收藏
err分享
err收藏
学者 查看更多内容