Improving the quality of reconstruction image by radon transformation by CNN. We will adpat the the Unet in Deep Convolutional Neural Network for Inverse Problems in Imaging[1] for our problems.
The convolution neural network is coded by tensorflow keras. Synthetic training dataset are generated by Matlab. It include 500 images of ellipses of random intensity, size, and location, sinograms for this data are created using matlab radon. After reproduction of the reference paper, we will consider to modify the CNN architecture to achieve image deconvolution for Terahertz CT imaging. Besides, we will investigate possiblity of realizing CT reconstruction by GANs rather than by traditional FBP or some other reconstruction algorithms. Our final target is realizing end-to-end Terahertz CT imaging reconstruction from blur convoluted sinograms to deconvoluted reconstruction images by deep neural network directly.
The Project will keep updating.
[1] Jin, Kyong Hwan, et al. "Deep convolutional neural network for inverse problems in imaging." IEEE Transactions on Image Processing 26.9 (2017): 4509-4522.