近幾年來QR碼在生活中已隨處可見，使用者只需使用手機相關應用程式即可馬上知道QR碼內所藏的資訊，因其便利性，有些商家、廣告商會將QR碼直接作為廣告文宣使用，將QR碼貼於平面及柱面上，為了吸引消費者的使用同時兼具設計感，部分商家會將QR碼的尺寸做得相較於一般QR碼大得許多，並將其黏貼於平面及柱面上，提供消費者拍照掃描，但是當QR碼貼附面積過大以至於無法在單一視角照到完整QR碼影像時(死角問題)，便無法順利解碼。
本論文我們將透過影像註冊的方法對多張片段資訊之QR影像進行拼接。首先我們需要使用者針對水平方向(左至右 or 右至左)對整張QR碼在不同視角盡可能得拍三張有部分重疊區域之影像，進而針對單張影像進行特徵點的偵測、篩選及過濾後，便能夠進行匹配。匹配時我們先透過交比 (cross ratio) [1]找出總模組數與QR碼單邊總像素值得比例關係，再搭配比對視窗的相關係數(cross correlation, CC)，便能夠進行初始的影像匹配，最後再透過隨機採樣一致性(RANdom SAmple Consensus, RANSAC)針對CC所配對出的匹配對進行篩選後，我們就可以找出影像間相互重疊的區域，接著使用投影轉換[3](projective transform)將重疊區域進行影像拼接，就可以得到一張近似於平面及柱面得完整QR影像供解碼器解碼。
最後我們模擬柱面及平面的影像註冊，由實驗結果得知，QR碼的錯誤更正級別(H、Q、L)對QR碼的解碼能力影響不大，當拼接的品質較為粗糙時，不論QR碼版本數的高低，皆會導致解碼失敗，相對的，當拼接的品質較為細膩時，成功解碼的機率便大幅提升。同時在本實驗當中，我們給予不同的參數與模組數搭配著CC及RANSAC的配對及篩選後，能夠減少影像註冊的時間，並將其重疊區域進行拼接，供解碼器解碼。

The QR code has become fairly common in our daily lives. The QR code allows us to acquire information easily via our smartphones. With such convenience, it has been adopted by a great number of stores and advertisers. Some even make their QR codes larger than usual so they can post them on both flat and curved surfaces, drawing passers-by’s attention. However, it can be difficult for a person to scan the QR code from a single angle(or blind angle problem) when it is overly magnified.
In this essay we conduct an experiment in which we put a few photos of a single QR code together through the image registration method. Doing so require a few steps. First of all, we take a few images of the QR code from three different horizontal perspectives. Secondly, we analyze and filter the feature points of the QR code in each image. Then, in preparation for matching the images, we figure out the relation between the total module number and one-side pixel of each image with the cross-ration method. Once acquiring the matching pair given by the CC, we utilize the RANSAC method to filter out the new matching pair from CC, then we can figure out the overlapping areas of the three images. Lastly, we put together the overlapping areas with the projective transform method. With such steps, we acquire a curved and flat image of the QR code.
After simulating image registration on both curved and flat images of the QR code, we figured that the error correction level (H, Q, L) of the QR code does not significantly affect its decodability. Lower merging quality leads to higher chance of failure in decoding, no matter what version the QR code is. On the other hand, greater merging quality results in higher chance of success. In addition, we acquired various parameters and module numbers with which we filtered and matched the feature points with CC and RANSAC, helping us tremendously to shorten the time for image registration.

[1]Brannan,Esplen,Gray. “Geometry, 5nd ed.” Cambridge Press, 2003.
[2]Z. A., M. B., M. M. E. V., and N. A. C. A., “CQR codes: Colored quick-response codes,” 2012 IEEE International Conference on Consumer Electronics - Berlin (ICCE-Berlin), pp. 321 - 325, 2010.
[3]Hartshorne, Robin. “Foundations of Projective Geometry, 2nd ed.” Ishi Press, 2009.
[4]Irwin Sobel, “History and Definition of the Sobel Operator” , 2014.
[5]Kalman, R. E. “A New Approach to Linear Filtering and Prediction Problems, “Transactions of the ASME - Journal of Basic Engineering Vol. 82: pp. 35-45, 1960.
[6]I. S. Reed and G. Solomon, “Polynomial Codes Over Certain Finite Fields,” Journal of the Society of Industrial and Applied Mathematics (SIAM), pp. 200 - 204, June 1960.
[7]My-Ha Le and Byung-Seok Woo and Kang-Hyun Jo, “A Comparison of SIFT and Harris Conner Features for Correspondence Points Matching”, University of Ulsan, 2011.
[8]M.A. Fischler and R.C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381–395, 1981.
[9]陳怡靜, “貼於平面及柱面上之帶有大標誌的QR碼之編解碼,” 國立台灣科技大學電子工程研究所,2015.
[10]周明豪, “透過透視投影對於黏貼於柱面QR影像進行校正及解碼,” 國立台灣科技大學電子工程研究所,2016.