Fast detection and segmentation of partial image blur based on discrete Walsh-Hadamard transform
Fast detection and segmentation of partial image blur based on discrete Walsh-Hadamard transform
复制标题
DOI：
10.1016/j.image.2018.09.007
复制
发表时间：
2019-02-01
影响因子：
3.5
通讯作者：
Zhou, Hongjun
中科院分区：
文献类型：
--
作者：
Wang, Xuewei;Liang, Xiao;Zhou, Hongjun
由于散焦或运动，图像信号可能会模糊。模糊可能不适合图像感知，但也可能包含有用的信息。因此，在机器视觉领域，检测自然图像中每个像素的模糊程度并分割部分模糊区域是一个重要而又具有挑战性的问题。本文提出了一种基于离散Walsh-Hadamard变换的简洁无参考方法来检测和分割局部模糊。首先，对从测试图像中提取的多个重叠图像块执行重新模糊策略。然后，对于测试图像和重新模糊的图像，在每个图像块上使用离散的Walsh-Hadamard变换来获得模糊映射。该模糊图可以表征测试图像中每个像素的模糊程度。在此基础上，结合K-Means聚类和区域生长，将测试图像分割成模糊/非模糊区域。在公共数据集上进行的实验证明了所提出的度量在模糊区域的检测和分割方面的能力。对比结果表明，该方法在散焦和运动模糊图像分割方面都具有优越性。该方法不需要数据训练，而且由于序列变换速度快，具有较高的时间效率。
Image signals can be blurred due to defocus or motion. Blur may be undesirable for image sensing, but may also contain useful information. Therefore, detecting the blurriness of each pixel and segmenting the partial blur regions in natural images are important and yet challenging in the field of machine vision. A concise no reference method based on discrete Walsh-Hadamard transform is proposed to detect and segment partial blur in this paper. First, a re-blurring strategy is performed over multiple overlapping image patches extracted from the test image. Then, for both test image and re-blurred image, discrete Walsh-Hadamard transforms are utilized in each image patches to obtain the blur map. This blur map can characterize the blurriness of each pixel in test image. Based on it, combined with K-Means clustering and region-growing, the test image can be segmented into blurry/non-blurry regions. The experiments, performed on a public dataset, demonstrate the capability of the proposed metric in the detection and segmentation of the blur region. Comparative results with the state-of-the-art show the superiority of the proposed approach in image segmentation for both defocus and motion blur images. The proposed approach is compendious without data training and possesses a high time efficiency because of the fast sequency transform.