报告（全英文）图像处理-图片找不同#Matlab下调用Python接口#SIFT算法应用.pdf-资料库

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第1页.png

第1页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第2页.png

第2页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第3页.png

第3页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第4页.png

第4页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第5页.png

第5页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第6页.png

第6页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第7页.png

第7页 / 共8页

359fae71-b276-4ff0-8a31-471c7ebd90b7.pdf-第8页.png

第8页 / 共8页

Final project of Image Analysis Technology and Applications Final project of Image Analysis Technology and Applications 1. Introduction The task of the project is finding the differences of the two images, ignoring butterfly. In addition, the code should be suitable for all image pairs. Given a butterfly template: And the image pairs: Fig.1 a butterfly template Fig.2 image pair 1 Fig.3 image pair 2 Fig.4 image pair 3 1

Final project of Image Analysis Technology and Applications To solve the problem, first, mark the butterfly's position. Second, align the two images. Finally, ignore the butterflies and find out what's different. Fig.5 image pair 4 2 Proposed approach The proposed approach includes marking butterflies(Matlab), image alignment (SIFT) (Python) , Due to the influence of the black butterfly stripe, binarization is very poor, so the first, I use filter. The maximum filter is selected here to remove the butterfly stripe. By extracting the edges, you can get the shape of a butterfly. find differences (Matlab) . 2.1 mark butterflies(Matlab) ①maximum filter ②extract the edge-canny ③region morphology imdilate\imfill\imopen ④Normalization of image size of butterfly pixel size. ⑤labeling After the normalization of image size, it is convenient to extract butterflies by utilizing the features Before the extraction of connected components, you need label the connected components. ⑥extraction of connected components-region features. The length and width of the template and the pixel feature are used to set the threshold for segmentation. 2.2 image alignment (SIFT) (Python) ①SIFT feature points and feature description extraction description. ②SIFT feature point matching SIFT algorithm is used to obtain the feature points in the image and the corresponding feature SIFT algorithm obtains the feature points in the image and the corresponding feature description. How to match the feature points in two images? In general, K nearest neighbor (KNN) algorithm can be used. The k-nearest neighbor algorithm takes the closest K data points in space and classifies them into a class. When matching feature points, KNN algorithm is generally used to find the two nearest neighbor data points. If the ratio of the closest and the second closest is greater than a given value, then we keep the closest value and consider that it matches the point as' good match '. ③Homography Matrix Through the above steps, we found a number of matching points in two graphs. How to align one graph with the other through rotation, transformation and other methods? 2

Final project of Image Analysis Technology and Applications And that's where the homology matrix comes in. Homography is the word for the same thing, Homo, which means the same thing, Homography, which means image, is the same thing that produces images. The points on the images from different perspectives have the following relations: Fig.6 The points relations Fig.7 the homography matrix As you can see, the homography matrix has eight parameters. If you want to solve these eight parameters, you need eight equations. Since each corresponding pixel point can generate two equations (x one, y one), you only need four pixels to solve this homography matrix. ④Random sample consensus:RANSAC SIFT algorithm has been used to find a number of matching points (dozens or even hundreds), so which four points should be selected to calculate Homography matrix? RANSAC algorithm cleverly solves this problem. It can effectively eliminate points with large errors, and these points are not included in the calculation of the model. RANSAC excludes some data points in each calculation, which plays the role of eliminating their noise, and calculates the model on this basis. Therefore, after obtaining a large number of matching points, RANSAC algorithm is used to select four random points at a time, and then the H matrix is obtained, and the iteration continues until the optimal H matrix is obtained. After OpenCV provides the above basic methods, we can use these methods to match images. Firstly, feature points in img1 and img2 are found, and then a better matching point pair is found. Finally, image img2 is projected and mapped by warpPerspective method. The calculation formula is as follows: Fig.8 perspective transformation Call the 'mark_butterfly' function to return the location of the butterfly. 2.3 find differences (Matlab) ①Matlab-mark_butterfly.m ②Python-cvsift.py ③Image subtraction Call the 'cvsift.py' to get the pixel-aligned image. going to lose a lot of information. So we subtract from each other and then add. ④Gray level transformation(imadjust) If you subtract two images, you're going to get a negative number, and if you just set it to 0, you're The effect of low gray value pixels was removed and the high gray value pixels were enhanced. About: imadjust f1=imadjust（f，[low_in high_in],[low_out high_out],gamma） 3

Final project of Image Analysis Technology and Applications Fig.9 imadjust transformation ⑤Median filter After gray scale transformation, there are still small noise points in the image, so the median filter Before the extraction of connected components, you need label the connected components . Remove the connected areas of butterfly position and mark different places. is used to process the image denoising. ⑥region morphology ⑦labeling ⑧extraction of connected components 3 Experimental results and analysis To validate maximum filter can remove the butterfly stripe. To validate canny can extract the edge. Fig.10 maximum filter To validate region morphology can get the shape of a butterfly. Fig.11 canny 4

Final project of Image Analysis Technology and Applications To validate bwareaopen function can get pixels in the area of 1,000 to 3000. Fig.12 region morphology To validate labeling can label connected component. Fig.13 bwareaopen To validate region features can extract of connected components. Fig.14 labeling To validate SIFT can align image. Fig.15 region features result 5

Final project of Image Analysis Technology and Applications To validate Call the 'mark_butterfly' function can return the location of the butterfly. Fig.16 image alignment To validate image subtraction can get the differences. Fig.17 the location of the butterfly To validate gray level transformation can remove gray value pixels and enhance the high gray Fig.18 image subtraction value pixels. To validate median filter can process the image denoising. Fig.19 gray level transformation 6

Final project of Image Analysis Technology and Applications To validate region morphology can get good connected components. Fig.20 median filter To validate labeling can label connected component. Fig.21 region morphology Remove the connected areas of butterfly position and mark different places. Fig.22 labeling Fig.23 mark different To validate the approach and the code can be suitable for all image pairs. 7

Final project of Image Analysis Technology and Applications Fig.24 image pair 1 result Fig.25 image pair 2 result Fig.26 image pair 3 result Fig.27 image pair 4 result 4 Conclusion For this experiment, we also encountered many difficulties and solved them one by one. Learned to use different tools for image processing. Also learned a lot of clever image processing methods. 8

资料库

报告（全英文）图像处理-图片找不同#Matlab下调用Python接口#SIFT算法应用.pdf

相关推荐

课程资源

热门标签

最新资料