SIFT与形态组件分析结合的无缝图像拼接方法

82 浏览量更新于2024-08-27 收藏 1.58MB PDF 举报

"基于SIFT和形态成分分析的图像镶嵌技术" 在传统的图像镶嵌过程中，显著的接缝问题一直是困扰研究人员的难点。为了改善这一情况，本文提出了一种新的方法，利用SIFT（尺度不变特征变换）算法和形态成分分析（MCA，Morphological Component Analysis）在训练词典上的应用，旨在使镶嵌结果更加平滑，提升视觉效果。图像镶嵌的核心在于图像配准和融合。首先，SIFT算法被用来提取图像的关键特征点。这些特征点具有尺度不变性和旋转不变性，能够有效地在不同的图像间进行匹配，从而实现图像的精确配准。SIFT算法通过检测图像中的尺度空间极值点，然后对这些点进行精确定位和描述，以确保在各种光照、缩放和旋转条件下都能准确识别。接着，我们利用K-SVD（Kernelized Singular Value Decomposition，核奇异值分解）算法对原始图像进行训练，构建一个过完备的词典。过完备字典能够捕捉到图像的复杂模式，为后续的分析提供更丰富的信息。K-SVD是一种用于学习字典的高效算法，它能够从数据中发现稀疏表示，对于图像处理任务特别有效。然后，MCA用于对图像进行分解，将图像分解为卡通部分和纹理部分。卡通部分通常包含图像的主要结构信息，而纹理部分则包含图像的局部细节。这种分离有助于在融合图像时分别处理这两部分，以达到更好的镶嵌效果。通过考虑这两个组成部分，图像的融合可以更加细致，从而减少接缝的可见性，提高整体的平滑度。此外，MCA的使用还有助于过滤原始图像中的噪声。由于噪声往往存在于局部细节中，MCA的分解过程能够识别并减少这些噪声，提高镶嵌图像的质量。结合SIFT算法的精准配准，K-SVD构建的过完备字典，以及MCA的图像分解与融合，本文提出的方法实现了先进的图像镶嵌性能，不仅提高了镶嵌结果的平滑度，同时也有效处理了图像中的噪声，提升了最终的视觉效果。这种方法对于遥感、全景摄影以及多视角图像处理等领域具有重要的实用价值。关键词：SIFT；图像配准；训练词典；K-SVD；形态成分分析

Image Mosaic Based on SIFT and Morphological

Component Analysis

Shuai Chen

, Yixiang Lu

1, ∗

, Qingwei Gao

School of Electrical Engineering and Automation,

Anhui University, Hefei 230601, China

Dong Sun

, Yi Xia

, Xueming Peng

1,2

Shanghai Huawei Technology

Co., Ltd, Shanghai 200120, China

Abstract—Due to traditional image mosaicing resulting in

remarkable seams, we propose a novel method to address

the problem to make the result more smooth. The approach

is constructed based on SIFT algorithm and morphological

component analysis(MCA) over trained dictionaries. The two

cores of image mosaic are image registration and fusion. First,

feature points are extracted by the SIFT algorithm which are

used to realize image registration. Then we employ K-SVD

algorithm to train an overcomplete dictionary based on the

original images and use the MCA to decompose the images so

that the decomposed components cartoon, texture can be used to

fuse images, respectively. This method leads to a state-of-the-art

mosaic performance, meanwhile, the noise contained by original

images can also be ﬁltered.

Index Terms—SIFT; image registration; trained dictionary; K-

SVD; MCA.

I. INTRODUCTION

Image mosaicing is one of the most important research

ﬁelds in image processing. Especially with the development

of computer vision and computer graphics, image stitching

technology is more and more mature and widely used in

various ﬁelds. Image mosaicing can be regarded as a process of

joining some images to construct a complete wide view image.

Image registration [1] and image fusion [2]are two main steps

in image mosaic process.

Image registration refers to determing the overlapping areas

of the image to be spliced, and then compute transformation

relationship. According to literatures, there are three meth-

ods are widely used in image registration: gray informa-

tion based(e.g., correlation methods [3]), transform domain

based(e.g., Fast Fourier transform based methods [4]), and

feature based methods(e.g, edges and corners etc). Among

them, the feature-based image stitching technology is popular

in image registration, because it can achieve higher matching

accuracy, time efﬁciency and robustness. Image fusion aims at

stitching the same scene of the two images taken at different

times, from different viewpoints, or different sensors in order

to form a smooth and seamless image. Image fusion can

be performed at pixel-level, feature-level, feature-level and

decision-level [5] [6]. At present, pixel-level fusion takes up a

large proportion in the fusion algorithms. Besides, pixel-level

fusion can also be implemented in spatial domain(e.g., weight-

ed averaged method, principal component analysis method

[7]) and transformation domain(e.g., discrete wavelet trans-

form(DWT) [8] [9] [10] and curvelet transform(CVT) [11]

[12]). However, these methods mentioned above consider the

whole image as a single component. And the information

of images can not be represented completely. So we took

the morphological component analysis [13] to decompose an

image into two components which made the presentation of

the image more complete and effective.

In this paper, we proposed a new method to mosaic im-

age which fuses the different components of the overlaps

independently in the sparse land. Compared with traditional

mosaicing methods, the knowledge of sparse representation

is introduced into image fusion, which can produce a more

smooth image effectively. Meanwhile, the noise contained

in the original images will be ﬁltered during the process

of dictionary training and image decomposition. The basic

process can be summarized as follows: ﬁrst, the SIFT [14]

algorithm is used to extract keypoints from the two images A

and B which capture different portions of the same scene with

an overlapped region. In fact, we use lightness to replace gray

level. According to the morphological component analysis, the

RGB color space is converted to HSV color space, which is

more suitable for color image processing. In fact, there also

exits an residual image for the noiseless image. Finally, the

cartoon, texture and residual images are fused respectively to

form new images(fused cartoon, texture and residual image),

and then construct a complete image by using the three images.

The fused image should be reconverted to RGB image. Fig.

1 shows the image stitching ﬂow based on point feature. In

summary, image mosaic consists of feature points extraction,

feature points matching, computing the transformation rela-

tionship and image fusion [15].

The rest of this paper is organized as follows: Section 2

gives a detailed description of the proposed method, Section

3 shows the experimental results of our novel methods, and a

conclusion is drawn in Section 4.

II. MATERIALS AND METHODS

A. Feature points extraction

The ﬁrst step in the image mosaic algorithm is to extract and

match features between two images. In practical application,

there exists many methods of feature extraction, such as Harris

corner detection [16], MSER (Maximally Stable Extremal

Region) [17] and the SIFT algorithm etc. In most cases, due to

there are variations of translation and rotation points between

two images in overlapping regions, therefore, in order to

2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics( CISP-BMEI 2017)

下载后可阅读完整内容，剩余5页未读，立即下载

weixin_38581405

粉丝: 2
资源: 947

SIFT与形态组件分析结合的无缝图像拼接方法

形态分量方法

利用MCA进行图像修复

Image Mosaic Method Based on SIFT Features of Line Segment

An improved image mosaic based on Canny edge and an 18-dimensional descriptor

基于Surf的Opencv全景图像拼接《Panoramic Image mosaic based on SURF algorithm using OpenCV》

sift-mosaic.rar_Sift mosaic_mosaic_sift 图像拼接_相位相关 MATLAB

image mosaic

Image Mosaic Using Speeded Up Robust Feature Detection:Image Mosaic Using Speeded Up Robust Feature Detection-matlab开发

ShowImage_mosaic.rar_C#马赛克_image mosaic_mosaic_图像马赛克

image_mosaic.rar_image mosaic_mosaic_全景图 matlab_图像拼接_多图像拼接

最新资源