2016/2/171 Image Vector Quantization Indices Recovery Using Lagrange Interpolation Source: IEEE International Conf. on Multimedia and Expo. Toronto, Canada, 2006 Author : Yung - Gi Wu and Chia - Hao Wu Student : Pei - Jun Jiang R Advisor : W. J. Chen
2016/2/172 Outline ► ► Introduction ► ► The proposed scheme ► ► Simulation results ► ► Conclusion
2016/2/173 Introduction ► ► Vector Quantization Developed by Gersho and Gray in 1980 Some other coding devised ► ► Discrete cosine transform, block truncation coding, wavelets coding, etc. One of the most successful signal processing techniques ► ► Quick and simple on decoding Encode ► ► Cut the image apart into the matrix of M *N ► ► Get a codebook, which each code vector has codewords ► ► Searches a best match code-vector in the codebook ► ► The best match code-vector have replaced the vector ► ► Transmit the code-vector index to the channel
2016/2/174 Introduction Decode ► ► Using these received indices to get code-vectors form the codebook ► ► Reconstruct the decoded image ► ► Internet transmission The indices may be lost ► ► Random noises ► ► Re-transmitting the data Wastes of time ► ► Estimate these lost data and to recover them Much more fast then re-transmitting the data Large data lost may cause the receiver determine the network disconnection
2016/2/175 Introduction ► ► Lagrange Interpolating Polynomial The first published by Waring in 1779 Rediscovered by Euler in 1783 Published by Lagrange in 1795 The polynomial P(x) of degree < or = (n-1) that passes through the n points (x 1,y 1 =f(x 1 )), (x 2,y 2 =f(x 2 )),..., (x n,y n =f(x n )),
2016/2/176 Introduction given by where Written explicitly
2016/2/177 Introduction ► ► The main ideal of the proposed method Decreasing the network traffic capacity Maintaining the network traffic capacity quality as well as possible Will not be considered when the data is lost seriously ► ► Causes network disconnection
2016/2/178 Outline ► ► Introduction ► ► The proposed scheme ► ► Simulation results ► ► Conclusion
2016/2/179 The proposed scheme ► ► Using Lagrange interpolation formula ► ► Two parts Preprocessing process Recovery process Fig 1 System diagram
2016/2/1710 The proposed scheme ► Preprocessing process Sort codebook ► Increase the relationship between near vectors of codebook Classify the code-vectors in the codebook as follow: Only sort the codebook once ► Before any other process ► Off-line Fig 2 Preprocessing process diagram D i : difference V : code-vectors i : i-th ε: threshold
2016/2/1711 The proposed scheme ► Recovery process Lagrange interpolation Uses the correct indices Set the lost index M,and L1, L2, R2, R1 are the correct indices ► Use L1, L2, R2, R1 to estimate and recover the lost M yiyi L1L1 L2L2 MR2R2 R1R1 xixi Fig 3 Recovery process diagram Table 1 Polynomial corresponding coordinates for each index
2016/2/1712 The proposed scheme ► ► Two-way polynomial Horizontal and vertical Different four coefficients for each way yiyi L1L1 L2L2 MR2R2 R1R1 horizontal[m-1,n-1][m,n-1][m,n][m,n+1][m+1,n+1] vertical[m-1,n+1][m,n+1][m,n][m,n-1][m+1,n-1] Table 2 Coordinates of coefficient L2L2 L1L1 MR1R1 R2R2 R2R2 R1R1 M L1L1 L2L2 Horizontal Vertical
2016/2/1713 Outline ► ► Introduction ► ► The proposed scheme ► ► Simulation results ► ► Conclusion
2016/2/1714 Simulation results ► ► Test image : Lenna,size is 512×512 codebook size is 256 threshold ε = 128 ► ► Random recovery Recovery those lose indices by random padding ► ► Low-pass filter Eliminate the random noise
2016/2/1715 Simulation results VQ indices lost-rate(%) Non recovery Random recovery Low-pass filter One-way Lagrange Two-way Lagrange Table 3 The quality of reconstructed image (Lenna) expressed in dB with different methods and different lost -rates
2016/2/1716 Simulation results ► In high data lost-rate Two-way Lagrange is powerful than other methods ► In low data lost-rate Two-way Lagrange is not so obvious Figure 4 The quality of reconstructed image (Lenna) with different methods and different lost-rates
2016/2/1717 Simulation results Figure 5 Original Lenna image Figure 6 The VQ reconstructed image (PSNR = 30.15dB) Figure 7 Lost-rate 1%Figure 8 Random indices recovery at lost-rate 1% PSNR = dB Figure 9 Low-pass filter recovery at lost-rate 1% PSNR =29.607dB Figure 10 Two-way Lagrange recovery at lost-rate 1% PSNR=29.735dB Figure 11 Lost-rate 5%Figure 12 Random indices recovery at lost-rate 5% PSNR = dB Figure 13 Low-pass filter recovery at lost-rate 5% PSNR =27.331dB Figure 14 Two-way Lagrange at lost-rate5% PSNR=28.418dB
2016/2/1718 Simulation results One hundred 512 X 512 gray level images Using two-way Lagrange to recovery Retransmission (T1 network :1.544Mbps) 31 milliseconds8290seconds Bit rate = 0.5 ► Recover one image is about 0.31 milliseconds ► Using two-way Lagrange to recovery Could be used in time requirement application Table 4 Compare the waste time of recovery images and retransmission images
2016/2/1719 Outline ► ► Introduction ► ► The proposed scheme ► ► Simulation results ► ► Conclusion
2016/2/1720 Conclusion ► ► An efficient data recovery method for VQ encoded image transmission ► ► If the data lost happens Do not request to transmit these data again ► ► Wastes of time Using the received correct data ► ► Estimate and recover the lost data ► ► Efficient in time-constraint situation ► ► Using Lagrange interpolating polynomial ► ► Property Fast processing The reconstructed images have visual acceptable quality
2016/2/1721 Thanks~