基于SIFT特征点的作物图像拼接算法研究

发布时间：2018-04-19 11:03

本文选题：图像拼接 + SIFT　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：基于特征的图像拼接技术因其能够更加充分的利用图像中的信息,使图像间的配准更加精确,已成为图像拼接技术中的主要研究方向。在特征的类型的选取上,尺度不变特征变换(Scale-invariant feature transform,简称SIFT)算法因其独特性好,产生数量多,鲁棒性强、可扩展性好等优点,而成为众多研究中主要选取的特征类型。本论文针对实验中使用传统SIFT算法进行作物图像拼接时,对图像非重叠区域部分提取特征点过多以及处理高分辨率作物图像时速度缓慢的问题,从拼接过程中的局部和整体这两方面各提出了一种新的改进算法策略,主要的研究内容如下。(1)从输入的6张农作物图像中检测SIFT特征点,每两张为一组利用基于k-d树的BBF算法进行特征点匹配,建立待拼接农作物图像间的匹配关系;然后,使用RANSAC算法选出正确匹配的特征点,计算出每组待拼接图像间的变换矩阵,并利用图像融合技术实现每组作物图像间的拼接;最后,将前两组的拼接结果与最后一组拼接,完成最终6张作物图像的拼接。通过编程实现及实验测试,针对所拍摄农作物图像中,主要景物间存在较为明显的像素对比度的特点,提出了一种适用于作物图像的快速特征点匹配方法。该方法通过提高图像中像素对比度的阈值,实现对图像中主要景物的特征提取,减少对图像中无用特征点的提取。实验表明,所提方法在处理农作物图像时,匹配精度相比较于使用传统的SIFT算法提升了10%左右,且速度也提升了1倍。(2)针对使用传统的SIFT算法进行农作物图像拼接时所存在的两个问题:(a)待拼接图像中非重复区域的特征点并不参与最终图像拼接的计算,且对于高分辨率的农作物图像来说更是增加了大量不必要的计算,占用了大量的内存,造成了时间上的严重浪费。(b)农作物图像中重复的景物,很容易在特征点匹配阶段形成错误的匹配点对,增加最终变换矩阵的计算时间。本文提出了一种处理高分辨率农作物图像的拼接策略,该策略先降低图像的分辨率,求出两幅待拼接图像间的大致重叠区域,再逐步对重叠区域求精,实现只对重叠区域的特征点进行检测,从而在一定程度上避免上述的两个问题,不仅提高了高分辨率农作物图像拼接的速度,而且提升了图像配准的精度。实验表明,该策略相比较于传统的SIFT算法在匹配精度上提升了20%左右。
[Abstract]:Feature based image mosaic technology has become the main research direction of image mosaic technology because it can make full use of the information in the image and make the registration between images more accurate.In the selection of feature types, Scale-invariant feature transform (sift) algorithm has become the main feature type in many researches because of its advantages of good uniqueness, large quantity, strong robustness and good expansibility.In this paper, we aim at the problem of too many feature points in the non-overlapping region and the slow speed of processing high-resolution crop images when we use the traditional SIFT algorithm in crop image stitching.A new improved algorithm strategy is proposed from the local and global aspects of the mosaic process. The main research contents are as follows: 1) detecting the SIFT feature points from the input 6 crop images.Each of the two pieces uses BBF algorithm based on k-d tree to match the feature points to establish the matching relationship between the crop images to be stitched. Then, the correct matching feature points are selected by using the RANSAC algorithm, and the transformation matrix between each group of images to be stitched is calculated.Finally, the results of the first two groups are spliced with the last group, and the final 6 crop images are stitched.Through programming and experimental testing, a fast feature point matching method is proposed for crop images, which has obvious pixel contrast among the main scenes.By raising the threshold of pixel contrast, the method can extract the features of the main scene in the image, and reduce the extraction of the useless feature points in the image.Experimental results show that the matching accuracy of the proposed method is about 10% higher than that of the traditional SIFT algorithm in crop image processing.And the speed is twice as high. 2) aiming at the two problems of crop image stitching using traditional SIFT algorithm, the feature points of the non-repeated region of the image to be stitched do not participate in the calculation of the final image mosaic.And for high-resolution crop images, it adds a lot of unnecessary calculation and takes up a lot of memory, resulting in a serious waste of time.It is easy to form the wrong matching point pairs in the feature point matching stage, which increases the computing time of the final transformation matrix.In this paper, a strategy of high resolution crop image stitching is proposed. This strategy firstly reduces the resolution of the image, finds out the approximate overlapping area between the two images to be stitched, and then refines the overlapped area step by step.In order to avoid the above two problems to a certain extent, it can not only improve the speed of high resolution crop image mosaic, but also improve the accuracy of image registration.Experiments show that compared with the traditional SIFT algorithm, the proposed algorithm can improve the matching accuracy by about 20%.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】