判别式模糊检测：提升图像清晰度的新方法

115 浏览量更新于2024-08-29 1 收藏 2.8MB PDF 举报

"判别式模糊检测功能：一种用于识别图像模糊和清晰区域的新方法" 本文主要探讨了在图像处理领域中一个重要的问题——如何有效地识别图像中的模糊与非模糊区域。作者通过研究图像梯度、傅立叶域以及数据驱动的局部滤波器中的模糊特征表示，提出了一种新的判别式模糊检测方法。这种方法与传统的基于恢复机制的方法不同，它旨在增强特征的判别力，并能适应图像中不同级别的模糊。在图像梯度分析中，作者可能研究了图像边缘的强度和连续性，因为模糊往往会导致边缘信息的损失或模糊。通过这种方式，他们能够检测出由于模糊而减弱的边缘，从而判断图像的模糊程度。在傅立叶域的研究中，作者可能利用了频域中的信息来识别模糊。图像模糊在频域中通常表现为高频成分的衰减，因此通过分析图像的频谱分布，可以识别出图像是否被模糊。数据驱动的局部滤波器则可能涉及深度学习或机器学习技术。这些方法能够从大量样本中学习到模糊和非模糊图像的模式，然后在未知图像上应用这些模式来进行分类。这种方法的优势在于它能够自适应地学习不同类型的模糊，提高检测的准确性和鲁棒性。为了验证所提方法的有效性，作者构建了一个新的模糊感知数据集，其中包含了数千张带有真实标签的图像。这个数据集的创建对于评估和比较不同模糊检测算法的性能至关重要，因为它提供了真实的模糊情况和相应的地面实况。应用方面，文章提到的几种应用包括： 1. 模糊区域分割：通过识别图像中的模糊区域，可以将其与其他清晰区域区分开来，这对于图像增强或编辑等任务非常有用。 2. 去模糊：检测到的模糊区域可以作为去模糊算法的输入，帮助恢复图像的细节。 3. 模糊放大：在保留图像原有模糊特性的同时，对模糊区域进行放大，可以提升低分辨率模糊图像的视觉效果。这篇论文提出的判别式模糊检测方法为图像处理领域提供了一种新思路，它能够更好地理解和处理图像模糊问题，有助于提升后续处理步骤的效果。这一工作对于图像处理、计算机视觉和机器学习等领域都有深远的影响，尤其是在图像恢复和增强技术上。

Discriminative Blur Detection Features

Jianping Shi

†

Li Xu

‡

Jiaya Jia

†

The Chinese University of Hong Kong

‡

Image & Visual Computing Lab, Lenovo R&T

jpshi@cse.cuhk.edu.hk xulihk@lenovo.com leojia@cse.cuhk.edu.hk

http://www.cse.cuhk.edu.hk/leojia/projects/dblurdetect/

Abstract

Ubiquitous image blur brings out a practically impor-

tant question – what are effective features to differentiate

between blurred and unblurred image regions. We address

it by studying a few blur feature representations in image

gradient, Fourier domain, and data-driven local ﬁlters. Un-

like previous methods, which are often based on restoration

mechanisms, our features are constructed to enhance dis-

criminative power and are adaptive to various blur scales

in images. To avail evaluation, we build a new blur percep-

tion dataset containing thousands of images with labeled

ground-truth. Our results are applied to several applica-

tions, including blur region segmentation, deblurring, and

blur magniﬁcation.

1. Introduction

Blur is one type of photo degradation that leads to loss of

details. In many special cases, it can also be a visual effect

purposely g enerated by photographers to give prominence

to foreground persons or other important objects based on

defocus or camera/object motion.

With the fast development of computer vision tech-

niques, it becomes important and practical to understand

information immersed in blurred images or regions. We ad-

dress a central blur detection problem in this area, since

quickly and effectively ﬁnding blur pixels can naturally

beneﬁt many applications including but not restricted to im-

age segmentation, object detection, scene classiﬁcation, im-

age quality assessment, image restoration, and photo editing

[6, 23, 21], given the fact that many blurred images exist on-

line or are produced from personal cameras.

There have been a series of methods directly solving

blind [4, 25, 7, 1 5] and non-blind [27, 12] deconvolution

problems. They aim at explicitly inferring latent images

and/or blur kernels. Our goal in blur detection is not to

follow this line using deconvolution [11]. Instead, we will

focus on ﬁnding and constructing blur feature representa-

tions directly from input images and making them potent

enough to differentiate between blurred and unblurred re-

gions, which are of high importance in feature understand-

ing.

A few previous methods relate to explicit blur detec-

tion. Levin [14] used image statistics to identify partial

motion blur. Lin et al. [16] also explored natural image

statistics for blur analy sis. Liu et al. [17] designed four lo-

cal blur features for blur conﬁdence and type classiﬁcation.

Chakrabarti et al. [3] analyzed direction al blur via local

Fourier transform. Dai and Wu [5] developed a two-layer

image model on a lpha channel to estimate partial b lur. Dif-

ferent from th e se approaches directly ﬁtting n atural image

statistics, we in this paper analyze feature discrep ancy in

gradient and Fourier space. We also propose a few features

that are with decent d iscrimination ability theoretically and

empirically.

In addition to feature construction, we explore a data-

driven solution, which learns local ﬁlters. We build a new

blur detection dataset that contains 1000 images with hu-

man labeled ground-truth blur regions. These data not only

make detection results convincing, but also provide useful

resource to understand blur with respect to structure diver-

sity in natural images. It enables training and testing, which

are traditionally hard to implement without suitable d ata.

Our contribution is three-fold. First, we design a set of

blur features in multiple domains. Second, we develop a

multi-scale solution for blur perception tha t avoids scale

ambiguity. Third, we build a blur detection dataset with

ground-truth labels on 1000 images, which provides a rea-

sonable evaluation platform for blur analysis. We apply our

results to several applications, including blur region seg-

mentation, image debluring and blur magniﬁcation.

2. Blur Features

We deal with challenging partially blurred images where

the point spread function (PSF) varies across the image.