Research and Comparison of Image Corner Detection Algorithm based on

Transcription

1 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun omputer Teaching Department, nyang Normal University, nyang Henan 455, hina bstract Image matching is a basic problem in image processing. Image corner detection plays a very important role in the ield o image matching and image understanding. The article describes three corner detection algorithm based on grayscale and uses # to program and compare the extraction results. The test results show that the Harris algorithm has a good detection rate and good stability in most scenarios, has practical signiicance or urther study..introduction Keywords: Image Matching, orner Detection, Harris lgorithm The image matching is a process aligning two images in space and determining the relative oset between the two grayscale images caught by two dierent sensors rom the same scene. The corner is a point whose D image brightness change violently [~]. The point is dierent rom the around points and has curvature maximum value on the edges o the image curve. [3] With the development o science and technology, image matching image inormation processing industry has become one o the key technology. The current most o the machine vision applications need image matching technology. Image matching is through the certain matching algorithm in two or more picture image recognition between the point o the process. Image matching is through the certain matching algorithm in two or more images o recognition between the same point, such as d image matching by comparing the target area and search area o the same size o the window o the correlation coeicient, take the search area o the largest correlation coeicient o the corresponding window center as the same place. Its essence is in the primitive similarity conditions, using the best matching criterion search problem. Image matching is mainly can be divided into gray based matching and signature-based matching. Gray matching principle: the image as a two dimensional signal, through the corresponding statistical method or signal between the matching relation, and then through the two signal correlation unction, analysis the similarity between them and get the point. Feature matching reers to two or more separately collection image eatures, the characteristic parameters, and then using description described parameters to match an algorithm. Image matching common eatures edge, linear, interest points, etc. ased on the characteristics o the image matching or image distortion has some robustness, matching perormance and eature extraction quality has strong relevance. Images o the characteristic inormation o the image processing is the oundation o the process. In the image characteristics o, the ngle point contains inormation data quantity minimum, also is the smaller data inormation storage the image gray scale luctuation characteristic inormation, and the intererence actor diagonal point eature extraction algorithm is less intererence. Thereore corner detection algorithm has an important application signiicance. Image matching technology has a wide range o applications, such as military target tracking and orientation, the medical image orming, geological prospecting, etc. Image matching algorithm adaptability, reliability, timeliness and positioning accuracy is relected matching algorithm o the main technical indicators. orner detection based on gray level o the assay, analysis pixels range o gray level change, calculated point curvature and gradient to detect image ngle point. Line eature is the image signiicant line eatures, such as buildings, road edge, etc. The traditional edge detection based on image line characteristics on the image o each pixel point or dierential and or second order dierential and then determine the edge pixels. ommon corner detection algorithm includes Harris algorithm, SUSN algorithm and MI algorithm. The Harris algorithm is based on a second-order directional derivative has good eect but Journal o onvergence Inormation Technology(JIT Volume8, Number6,Mar 3 doi:.456/jcit.vol8.issue6.6

2 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun large amount o computation. SUSN [4] algorithm has higher eiciency and strong anti-noise ability. MI algorithm has high precision and good stability..orner detection algorithms.. Harris algorithm Harris algorithm has a high stability and robustness. It can accurately extract the corner in the rotating images or gray changing images which is suitable or the extraction o multi-sensors image corner eature. The Moravec algorithm is the predecessor o the Harris algorithm. The ollowing igures give the correlation window o our dierent positions. (a (b (c (d Figure The correlation window o our dierent positions Fig. is our dierent positions: (a represents that the current pixel is located inside the object on the assumption that the gray level is invariant in any case such as moving windows, (b indicates that the pixel is on the edge o the object, the gray level change greatly i move associated window along the vertical edge, (c represents the pixel is corresponding to a corner, the gray level change slightly i move along edges, (d indicates the pixel is corresponding to a isolated point, the gray value change greatly no matter how to move the associated window. [5~6] The corner response unction o Harris algorithm is deined as: R Det( M ktr( M ( orner response criterion R is positive in the corner region, negative on the edge area and very small in the unchanging area [7~8]. In the practical application, the point is a corner point i R greater than a given threshold value. The limitation o Harris algorithm is that it use the higher order partial derivative method, thereore it is sensitive to the noise and has high computational complexity when carry out Gaussian convolution operation. The gradient has anisotropic because o the only considering o the horizontal and vertical directions when calculated...susn algorithm SUSN algorithm is based on Univalue Segment ssimilating Nucleus. SUSN algorithm uses a circular template (also known as the window or nuclear with isotropic eature or corner detection. The general template radius is 3-4 pixels. Fig. is a 7*7 template with a ixed threshold g=37 /.

3 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun Figure. 7 7 template The luminance values o the points in the template are compared with that o the nuclear point (current point by the ollowing similarity unction. Here we use the template with 37 pixels. c r r,,, I r I r t I r I r t ( c is an output unction o comparison calculated by the pixels in the template. In the equation, r represents the location o the nuclear point o the D image, r is the location o other points, I( r is the luminance value o a pixel in the template, t is a threshold o luminance dierence. USN can be calculated by the ollowing equation: n ( x, y c( x, y (3 ( x, y ( x, y n represents the number o pixels in area USN. omparing n with a special threshold value g, SUSN corner detection algorithm will set g as hal o the n max (the maximum area o n max. The initial response unction can be obtained through the ollowing equation. R r, nr g nr nr g, g That is a simple SUSN algorithm ormula. g deines the maximum value o USN o the output corner point. Threshold t in ormula (4 represents the minimal contrast ratio o the corner point detected, which is the maximum tolerance to noise. The value o t determines the quantity o the eature extracted. The smaller t is, the more eature extracted rom the image with low contrast ratio. Thereore, or dierent contrast ratio and dierent noise image the t values are also dierent. In actual applications, the stable comparison unction shown as ollow is oten adopted: (4 r r r r I( I( c(, exp( t 6 (5 SUSN corner detection algorithm can be described as: long pixels be detected, in accordance with the order traversal each point on the image. When reach to a pixel, this pixel point is regarded as the core. Then scan along the arc with radius o R to identiy the point on the arc where the gray-scale variation strongly.

4 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun.3.mi algorithm MI uses the deinition o USN. Suppose that any line cross the nuclear point has two intersections with the circle window P and P'. Deine the RF (orner Response Function as ollow: is the nuclear point, p grey level o point P. R min(( ( (6 P P' ' Figure3. the linear dierence diagram The detection principle o MI algorithm is to ind the minimal brightness variation R in special circular window. I R lager than the threshold, it is the corner, else isn't. We usually use a nearly circular discrete window in the calculation. min (7 R (( P ( P' ' PP, ' Sn is the nuclear point. Point P and point P ' are symmetrical on RF calculation is divided into two steps. First the change o brightness along the horizontal and vertical should be calculated: r r ( ( ( ( ' ' (8 The value o the corner response unction is : R min( r, r (9 linear internal pixel interpolation method is used to calculate the change value o minimum luminance: RF R min( r ( x, r ( x ( Parameter x is deined as the intersections o any line across N and the circle window, x (,. Usually we take the position o the intersection o square side.

5 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun The reaction unction is: r( x ( ( P P' r ( x ( ( Q Q' ( Since the circular symmetry, the equation can be simpliied as: P P' Q Q' ( x ( x ( x ( x ' ' x x x x ' ' ( r. Take ( into (: ( r ( r, r ( r ( r r ( x x r ( x x x x (3 In the equation: r r r ( ( r r ( ' ( ' N ( N ' ( ' ' ( ' ( N N (4 To ensure the RF value the largest in the minimum value, we get >. orner response unction has minimum value on the square, so <, <, +>, then we obtain the value o RF as: R / (5 I this value doesn't meet the requirement, the value will be calculated by equation (9. 3. System implementation Development environment Here we use #. # is a development tool o.net ramework, a modern & object-oriented programming language. System Framework System interace is mainly divided into two parts: basic operation and corner detection. System unction module chart is shown as ollow:

6 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun Figure4. System unction module chart 3Harris algorithm realization: Suppose the gray level o pixel whose coordinate is (x,y in the image I(x, y,pointer is I[x, y]. Each pixel o the image is iltered by horizontal and vertical dierence operator to get Ix, Iy : I * Ix I * Iy (6 Ix, Iy and Ix* Iy can be calculated 3 The 5*5 D Guass template with zero mean is set as w w ( i ( j guassi, jexp * sigma (7 Normalize the template, inal Guass template is obtained. 4 Gaussian ilter the three elements, that is use Guass template convolute with Ix, Iy, Ix* Iy. The characteristic values or the correlation matrix M o each pixel can be calculated through these values. M Ix * Iy Iy Ix Ix* Iy (8 5 The corner response unction cim is calculated: cim i, j Ix Iy Ixy K( Ix Iy (9 The value o K is between.4 and.6, here K equals to.5. 6 Seeking local maximum to determine the inal corner. In matrix cim, i cim is the local maximum value and larger than thresh, the point is considered the corner.

7 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun 4ode implementation n array is deined to store binary data, and its code is shown below: public ircle( public int[,] imbw(itmap image { int [,] bw=new int [image.height,image.width]; olor =new olor (; int r=; or(int a=;a<image.height;a++ or(int b=;b<image.width;b++ { =image.getpixel(b,a; r=.r; bw[a,b]=r; } return bw; } 4.Experiment analyses Figure5 is the standard image commonly used in the ield o corner detection, which contains a large number o distinct geometry corner inormation while the transition regions on the let side contains unobvious corner inormation. Figure6 is a plus-noise eect diagram used to detect the noise resistance o dierent algorithms. In order to compare dierent algorithms or dierent image processing, the experiment also used building blocks diagram Fig.7 and Housing diagram Fig.8 or comparison. Figure5. Standard diagram Figure6. plus-noise diagram Figure7. building blocks diagram Figure8. Housing diagram ompare the eect o dierent parameters to the three algorithms in Fig.4. ( Harris corner detection algorithm

8 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun The three parameters mentioned above have some eect in Harris corner detection. The other data is in Table 4.: (a Threshold 5 (b Threshold 5 Figure9. Harris changing threshold test 3 Table. Harris changing parameter test Threshold orner number Guass window orner number Gaussian variance.7.8 orner number ( SUSN corner detection algorithm The related data is in Table : Table. SUSN changing parameter test Threshold t orner number Threshold g orner number (3 MI corner detection algorithm The related data is in Table 3: Table 3. MI changing parameter test threshold orner number The comparison o detection time and corner number o three algorithms, the results are shown in Table 4

9 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun Figure number lgorithm Table 4. Result statistics Detection time verage time (ms (ms orner number verage number Harris Fig SUSN MI Harris Fig SUSN MI Harris Fig SUSN MI With high computational complexity, Harris algorithm is sensitive to noise. The inluence o noise can be reduced by increase the Gauss window. Harris algorithm is anisotropic whose perormance is strong reaction to the corners making it's easy to error detection. SUSN algorithm has strong anti noise capability and high accuracy which can be used to detect all types o corners. ut the calculation o SUSN algorithm is complex, when the contrast level is not the same in dierent area, the result under ixed threshold t oten do not meet the actual situation, and extraction eect is not evident in the gradient region. MI corner detection algorithm has high accuracy, good stability and ast operation. ut it does not have rotation invariance and sensitivity to noise, the edge point can be erroneously detected as a corner, these actors aect the stability o the algorithm. When using a large template operator, a lot o corners will be missed and corner location is inaccurate.

10 Research and omparison o Image orner Detection lgorithm based on # JIN Xian-hua, M xiao-jun 5. onclusions Overall, Harris algorithm is not aected by light & rotation, although with low positioning accuracy and high computational complexity, it has a good detection rate & anti-noise perormance and good stability in most cases, so it is the optimal algorithms in this article. For many applications, the positioning accuracy is not the most important, at the same time, due to the rapid development o computer hardware and increasing computational eiciency, the Harris algorithm has been used most widely. 6. Reerences [] Yu-Long Qiao, Fu-Shan Wang, hun-yan Song, "Wavelet-based Dynamic Texture lassiication Using omplex Generalized Gaussian Distribution", ISS, Vol. 4, No. 4, pp. 9 ~ 7, [] JianHua du,, MingHui Wang, ZhenYa Wu, Jun Hu, "Inrared Image nd Visible Light Image Fusion ased On Nonsubsampled ontourlet Transorm nd The Gradient O Uniormity", IJT, Vol. 4, No. 5, pp. 4 ~, [3] Hu Wang, Wensheng Zhang, Jin Liu, "Description or 3D Images with Histogram o Gradients ased on Maximum Directional Derivative", JDT, Vol. 6, No. 4, pp. 5 ~ 3, [4] Ken ai, Xiaoying Liang, Rong-qian Yang, ShanXin Ou, "utomatic Whole Heart Segmentation o 4D Dual Source ardiac T ased on the Deormation Field", JDT, Vol. 6, No. 8, pp. 36 ~ 369, [5] yung-ki Hong, Seung-Joong Shin, Jae-Woong hoi, and Kae-Dal Kwack, "Error Resilient Interlace to Progressive onversion lgorithm or Noisy Image", RNIS, Volume, pp. 5 ~, 8 [6] hiew Kang Leng, Jane Labadin, Sarah Flora Samson Juan, "Steganography: DT oeicients Reparation Technique in JPEG Image", JDT, Vol., No., pp. 35 ~ 4, 8 [7] N. Krishnan, P. Vijayalakshmi, R.K. Selvakumar, S.Naga nandini sujatha, "daptive Source pixel based Prediction or Lossless Intra oding o H.64 MPEG-4 /V", JIT, Vol. 3, No., pp. pp.57 ~ pp.63, 8 [8] Seyed Zeinolabedin Moussavi, Saeedreza Ehteram, li Sadr, "The New Face Recognition Technique With the use o P and LD", JIT, Vol. 3, No., pp. pp.34 ~ pp.38, 8