22.08.2015 Views

Jian-Jiun Ding, Wei-Lun Chao, Jiun-De Huang, and Cheng-Jin Kuo

Jian-Jiun Ding, Wei-Lun Chao, Jiun-De Huang, and Cheng-Jin Kuo

Jian-Jiun Ding, Wei-Lun Chao, Jiun-De Huang, and Cheng-Jin Kuo

SHOW MORE
SHOW LESS
  • No tags were found...

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5yProposed Fourierdescriptor schemeSnAnti-symmetric2N 3kConventionalFourierdescriptorx SNS 0 1 0BoundarycompressionBoundaryPre-segmentationSeg-FDReconstructionwith CR = 2.5%ypreserving property of the proposed Fourier descriptor scheme:No matter how many DFT coefficients are remained, the endpoints of the reconstructed segment will coincide with those ofthe original segment. This proof can be achieved by checkings W *(0) <strong>and</strong> s W *(N-1), which are the end point locations of thereconstructed segment of s W (n). As mentioned in (2), the twoend values of s W (n) are 0; then if s W *(0) <strong>and</strong> s W *(N-1) are also 0,the reconstructed non-closed segment s*(n) after the inverseoperations of (2) will have the same end points as what s(n) has.The formulas of s W *(0) <strong>and</strong> s W *(N-1) are presented as follows:P2N3j20 k/(2N2) j20 k/(2N2) k1 k2N2PP2N3s *(0) S( k) e S( k)eWk1 k2N2PPk 1Sk ( ) Sk ( ) Sk ( ) S(2N2 k) 0,(8)P2N3j2 ( N1) k/(2N2) j2 ( N1) k/(2N2)k1 k2N2PP2N3jkSke( ) jkSke ( )k1 k2N2PPSk ( ) S(2N 2jkk) e0k 1s *( N 1) S( k) e S( k)eWxReconstruction(Step 2 & 3)S*n P coefficients +2 end point locationsReconstruction(Step 1)Anti-symmetric2N 3kZero paddingFig. 6. Illustrations of both the boundary compression <strong>and</strong> the reconstructionprocesses: After applying the proposed scheme, P of the coefficients <strong>and</strong> thetwo end point locations are remained for boundary recording. To reconstructthe non-closed segment, we first regenerate the (2N-2)-long coefficientsequence (Step 1) <strong>and</strong> then perform Steps 2 & 3 in Section III-D to achieve it.where the anti-symmetric property of S(k) is considered, whichis derived from the proposed anti-symmetric extension; withoutanti-symmetric extension, the end point preserving property isnot guaranteed after boundary compression. Note that both (8)<strong>and</strong> (9) are held under arbitrary P (0≦ P ≦ N-2).(9)ProposedFourierdescriptorFig. 7. The flowchart <strong>and</strong> illustration of the proposed Seg-FD framework forclosed boundary: The top row shows the reconstruction result using theconventional Fourier descriptor [4] with 2.5% compression rate. The bottompart shows the result of Seg-FD (the dash box). As shown, the proposedSeg-FD framework could lead to better reconstruction quality, especiallyaround the corners.IV. THE PROPOSED METHOD FOR CLOSED BOUNDARIESThe proposed Fourier descriptor scheme could directly beapplied on closed boundaries; the resulting Fourier descriptors<strong>and</strong> reconstructed boundaries are similar to the ones achievedby applying the convectional Fourier descriptor [3]-[4], [15] —since the linear offset step mentioned in Section III-B does notchange the shapes of the originally closed boundaries.Generally, Fourier descriptors are very efficient <strong>and</strong>effective for recording closed boundaries; under highcompression rates, however, detailed structures of the originalboundaries such as corners <strong>and</strong> sharp angles still cannot bemaintained. To alleviate this problem, we further propose aframework named Seg-FD:Seg-FD performs boundary Segmentation at first togenerate several non-closed yet smooth segments, <strong>and</strong> thenapplies the proposed Fourier <strong>De</strong>scriptor scheme to record thesesegments. The flowchart of Seg-FD is depicted in Fig. 7.A. Boundary Pre-segmentationIn order to create smooth non-closed segments, sharp anglesor corners on the closed boundary should be detected first; theclosed boundary is then divided at these positions into severalnon-closed segments containing no corners inside. There havebeen many corner detection algorithms in existing literatures,like the Harris corner detector [6], SIFT [9], <strong>and</strong> the machinelearning approach in [12]. These methods, nevertheless, aredesigned mainly for corner detection in regular images, <strong>and</strong>either with high computational complexity or requiring alearning step. For our cases on binary contour images <strong>and</strong> forefficient detection, a simple algorithm through checking theinner products along contours (shown in Fig. 8) is utilized.Given an N-point boundary {(x n , y n )| n = 0, 1, …, N1}thiscorner detection algorithm goes through each point (x n , y n ) in

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!