Vol13 (Softech 016), pp138-14 http://dxdoiorg/101457/astl016137 Video Seamless Splicing Method Based on SURF Algorithm and Harris Corner Points Detection Dong Jing 1, Chen Dong, Jiang Shuen 3 1 College of Information and Electrical Engineering, China Agriculture Universit Beijing, 100083, China School of Information Science and echnolog, Beijing Forestr Universit Beijing 100083, China 3 Beijing Research Center for Information echnolog in Agriculture Beijing 100097, China Abstract In order to improve image registration accurac and efficienc and solve the problems existing in single Harris corner point detection algorithm or SURF algorithm as ell as take full advantages of the to algorithms, an image registration algorithm based on corner point detection and feature point integration is proposed in the article he simulation result shos: compared ith other image registration algorithms, the algorithm proposed in the article not onl improves image registration accurac and speed, but also has strong robustness, thus having ide application prospect in image analsis field Ke ords: Harris corner points; SURF feature; neutral netork; image registration 1 Introduction As image application basis, image processing includes image segmentation, image registration, etc, herein the image registration refers to the processing procedure of establishing the corresponding relation for to or more images obtained at different time and from different angles or different sensors, and has ide application prospect in such fields as motion estimation, remote sensing, medical imaging and compute vision, so it is alas the research hotspot in image processing field [1] Harris algorithm Harris algorithm emplos differential operation and autocorrelation matrix to detect corner points and has the features of simple calculation, uniform and rational features of extracted corner points, quantitative extraction of feature points and stable operator he process of carring out Harris corner point detection for gre edge image I g to generate corner point image I c [7] can be expressed as follos: ISSN: 87-133 ASL Copright 016 SERSC
Vol13 (Softech 016) Gaussian indo function W(u, v) is used to calculate image partial derivative, and 3*3 Gaussian indo function is selected in the article and (δ x, δ ) is used to convert indo function W(u, v), namel: E(, ) ( ) (, ) x I u x I u v (1) alor series expansion is carried out for the first item of formula (1), namel: Ix( u x, v ) I( u, v) Ix( u, v) I ( u, v) x, () herein, I x and I respectivel denote partial derivatives of x and directions; formula () is put in formula (1) to obtain the folloing formla: E( x, ) I x( u, v) I ( u, v) x, x, M x, (3) In the formula, M is autocorrelation matrix and the calculation formula of M matrix can be obtained from formula (3), namel: M ( I x( u, v)) I x( u, v) I ( u, v) I ( u, v) I ( u, v) ( I ( u, v)) x (4) If M has to small characteristic values, then it is indicated that the present point is located at flat area; if M has one large characteristic value and one small characteristic value, then it is indicated that the present point is located at the edge If M has to large characteristic values, then it is indicated that the present point is corner point Harris has also provided another formula to evaluate hether this point is corner point or not R DE M K trace M ( ) ( ) (5) In the formula, R denotes corner point value; k is an adjustable sensitive parameters 3 Particle sarm optimized neural netork BP neural netork is a multilaer feed-forard netork, herein the input signal is transmitted to output laer from hidden unit and the output signal is generated at the output end, and this is actuall the forard transmission of the orking signal he error beteen the actual output and the expected output of the netork is regarded as the error signal hich is transmitted forards from the output end laer b laer and Copright 016 SERSC 139
Vol13 (Softech 016) continuousl corrects netork eights and threshold value during the backard transmission so as to make the actual output of the netork closer to the target output Additionall, the topological structure of BP neural netork model includes input laer, hidden laer and output laer, as shon in Fig 4 [9] x 1 x x j ij jk 1 k Input Laer Hidden Laer Output Laer Fig 1 Basic Structure of BP Neural Netork 4 Simulation environment In order to verif the performance of the image registration algorithm in the article, a simulation experiment is carried out on the platform ith the hardare environment as: 30GHzCPU, GB RAM, Windos XP operating sstem and the softare environment as: MALAB 01 softare As the simulation object, Einstein s image is respectivel moved b (937, 815), (045, 184) and (954, 865) and is added ith impulse noise ith the noise densit of 01, as shon in Fig 6~8 (a) Reference Image (a) Moved Image Fig 6 Movement (937, 815) 140 Copright 016 SERSC
Vol13 (Softech 016) Fig Movement (045, 184) (a) Reference Image (a) Moved Image (a) Reference Image (a) Moved Image Fig 3 Movement (954, 865) and Noise Addition (Densit: 01) 5 Conclusion In order to improve video image registration accurac during splicing, a video image registration algorithm based on corner point detection and feature point integration is proposed in the article Firstl, Harris algorithm is adopted to detect image corner points and fetch surf feature points of the image as ell as integrate the corner points and feature points and eliminate the repeated points to generate feature vector; secondl, neural netork is used to learn the image vector and establish image registration model, and the particle sarm algorithm is used to optimize the parameters of neural netork; finall, the specific image is adopted for simulation comparison experiment he result shos: compared ith other image registration algorithms, the algorithm proposed in the article can improve image registration accurac and efficienc and has relativel strong robustness References 1 Lv, Z, Feng, L, Feng, S, Li, H: Extending ouch-less Interaction on Vision Based Wearable Device Virtual Realit (VR), 015 ieee IEEE, 015 Zhang, M, Lv, Z, Zhang, X, Chen, G, Zhang, K: Research and Application of the 3D Virtual Communit Based on WEBVR and RIA Computer and Information Science, no 1 (009): p84 Copright 016 SERSC 141
Vol13 (Softech 016) 3 He, J, Geng, Y, Pahlavan, K: Modeling Indoor OA Ranging Error for Bod Mounted Sensors, 01 IEEE 3nd International Smposium on Personal Indoor and Mobile Radio Communications (PIMRC), Sdne, Australia September 01 pp68-686 4 Geng, Y, Chen, J, Pahlavan, K: Motion detection using RF signals for the first responder in emergenc operations: A PHASER project [C], 013 IEEE 4nd International Smposium on Personal Indoor and Mobile Radio Communications (PIMRC), London, Britain September 013 14 Copright 016 SERSC