基于改进FREAK算法的图像拼接技术研究

发布时间：2018-03-31 23:03

本文选题：图像拼接　切入点：图像匹配　出处：《哈尔滨工程大学》2016年硕士论文

【摘要】：图像拼接作为计算机视觉方面的一门关键学科,在许多领域和日常生活范畴内都有着普遍的应用。图像拼接技术的产生就是为了解决宽视角与高空间分辨率不能兼得的缺陷,这成为大量科研工作者的研究热点,使普通的成像设备和一般的数码相机都能获取高清晰度的宽视野图像。图像拼接技术是针对部分重叠的图像序列,在不降低原图像清晰度的情况下使用计算机中的一些图像处理算法,对图像进行特征信息的配准,再利用融合算法使拼接后的图像具有宽视野、高清晰度、无缝等特点。总体上可理解为,图像拼接的过程主要涵盖重要两步:图像配准和图像融合。本文首先叙述了关于图像拼接的基础概念和相关技术,有关图像配准方面阐述了几种传统常用的特征点匹配算法,经过对比后发现传统算法对图像间发生尺度变化的匹配效果不好的劣势,提出了一种基于改进FREAK算法的特征点匹配算法。改进的算法将SURF算法和FREAK算法有效联系为整体用于特征点检测与筛选:利用SURF算法中Hessian矩阵确定候选点,进行非极大值抑制,建立尺度空间;然后用FREAK算法描述得到的特征点,并分配方向,最终得到图像匹配结果。实验表明,改进算法在图像特征点匹配准确度上有一定的提高,并且对图像的尺度差异、光照差异以及旋转差异具有良好匹配效果。对给定部分重叠的原图像实行拼接,使用本文提出的改进FREAK算法应用于原图像实行特征点检测与筛选,有利于图像间的对齐准确率;再结合RANSAC算法提纯匹配对,进一步分析建立单应性矩阵在新坐标系下统一图像的视角;接着采取线性渐变加权算法消除配准产生的接缝,从而达到图像拼接技术的实际应用,经处理后图像具有宽视野、高质量的特点,具有良好的视觉效果。针对图像融合技术提出了以小波变换为基础的改进的融合准则方法,具体方法为小波变换把图像解析为高频系数与低频系数,高频系数采用与两种Laplacian模板算子卷积,并求其图像熵函数作为清晰度评价判断函数处理,再进一步对熵函数比较求取最终系数值;低频系数借鉴Otsu算法思想利用阈值结合拉普拉斯清晰度评价函数的判断规则处理,将原图像的能量最大限度提取,尽可能保留原图像的概貌,最终使用小波逆变换重构显示出融合效果。对于融合结果图,通过主观图像质量评价并结合客观图像质量评价的方法说明改进算法的有效性。
[Abstract]:As a key subject in computer vision, image mosaic has been widely used in many fields and daily life. This has become the research hotspot of a large number of researchers, so that ordinary imaging equipment and general digital cameras can obtain high-definition wide-field images. Image stitching is aimed at partially overlapping image sequences. Some image processing algorithms in the computer are used to register the feature information of the image without reducing the clarity of the original image, and then the fusion algorithm is used to make the stitched image have wide field of vision and high definition. Generally speaking, the process of image stitching mainly covers two important steps: image registration and image fusion. On the aspect of image registration, several traditional feature point matching algorithms are described. After comparison, it is found that the traditional algorithm has poor matching effect on the scale change between images. A feature point matching algorithm based on improved FREAK algorithm is proposed. The improved algorithm effectively links the SURF algorithm and the FREAK algorithm as a whole for feature point detection and screening. The candidate points are determined by using the Hessian matrix in the SURF algorithm and the non-maximum suppression is carried out. The scale space is established, then the feature points are described by FREAK algorithm, and the direction is assigned. Finally, the result of image matching is obtained. The experimental results show that the improved algorithm can improve the accuracy of image feature points matching to a certain extent. And it has a good matching effect on the scale difference, illumination difference and rotation difference of the image. The original image with a given overlap is stitched, and the improved FREAK algorithm proposed in this paper is used to detect and filter the feature points of the original image. It is beneficial to the accuracy of image alignment, and then combines the RANSAC algorithm to purify matching pairs, and further analyzes the establishment of monogram matrix in the new coordinate system unified image perspective; then the linear gradient weighting algorithm is adopted to eliminate the joint produced by registration. In order to achieve the practical application of image mosaic technology, the processed image has the characteristics of wide field of vision, high quality and good visual effect. An improved fusion criterion method based on wavelet transform is proposed for image fusion technology. The specific method is that the image is analyzed into high frequency coefficient and low frequency coefficient by wavelet transform, and high frequency coefficient is convolution with two kinds of Laplacian template operators, and its image entropy function is used as the definition evaluation function. Furthermore, the final coefficient is obtained by comparing the entropy function, and the low frequency coefficient is extracted from the maximum energy of the original image by using the threshold value combined with the judgment rule of Laplacian definition evaluation function for reference of Otsu algorithm. Finally, the fusion effect is shown by inverse wavelet transform. The effectiveness of the improved algorithm is illustrated by subjective image quality evaluation and objective image quality evaluation.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】