ECE 539 Intro-ANN Gaoang Wang Image Reconstruction based on Back-Propagation Learning in Compressed Sensing theory ECE 539 Intro-ANN Gaoang Wang
1. Compressed Sensing Theory Since y=Φx, x=Ψf, then y=Φx=ΦΨf Where x is original signal, y is sampled signal, Φ (m by n) is measurement matrix with m<<n, and Ψ is transform matrix UW-Madison
2. SPL-BCS (by James E. Fowler) UW-Madison
3. Learning Method in Sampling Feature vector (64 by 1): from one patch (8 by 8) of sampled image. pinv(mea_matrix)*mea_matrix*patch 5 Input Images (256 by 256) have 5120 patches in total as training data. Using back-propagation learning method to derive the weights of classification. UW-Madison
4. Modified Sampling Method First sampling: sampling input images in general and using pre-generating weights to decide which part of the original image can bear a higher compression ratio. Second sampling: transform the satisfied patches into lower dimension. UW-Madison
5. results Image C_Rate C_Ratio PSNR BCS Ratio Lena 0.8572 14.2510 28.5831 8 27.2254 Barbara 0.7781 12.8037 23.5672 23.3518 Goldhill 0.7739 14.2598 27.1093 24.1568 Baboon 0.6860 11.2598 21.1042 20.3121 Peppers 0.8831 14.0137 28.6339 27.7714 test1 0.8921 14.4414 29.7981 29.6381 test2 0.8701 14.4590 27.1122 26.8491 test3 0.7905 14.1924 26.2250 26.4124 test4 0.8435 14.8369 28.2777 28.0461 UW-Madison
5. results UW-Madison
Thank you ! UW-Madison