Sunday, June 25, 2017

Two Recent Theses

University of Michigan publishes PhD thesis "CMOS Sensors for Time-Resolved Active Imaging" by Jihyun Cho. After a quick overview of image sensor architectures and noise, the thesis describes a single shot FLIM imager and a ToF imager with background light suppression:


Lund and Linköping Universities, Sweden, publish MSc thesis "Demosaicing using a Convolutional Neural Network approach" by Karin Dammer and Ronja Grosz. The main intention of CNN approach is reduction of artifacts:


The results are somewhat mixed: "Using convolutional neural networks is a valid method for demosaicing images with good results and it could replace a method using linear interpolation. Our CNN method outperforms the multilayer perceptron by a difference of 7.14 dB in the peak signal to noise ratio. The convolutional neural network performs well when using L2 and PSNR as loss functions when training the network, however SSIM does not perform as well. Despite the relatively good result the network would benefit from using an error metric that is better at indicating the presence of image artifacts and color errors. The network did not significantly benefit from the residual layer nor a deconvolution layer."

No comments:

Post a Comment

All comments are moderated to avoid spam and personal attacks.