דוד בובליל (הנדסת חשמל, טכניון)
יום שלישי, 31.10.2017, 11:30
חדר 337, בניין טאוב למדעי המחשב
In the past decade we are experiencing a massive come-back of Neural Networks that flourished in the eighties, this time with a much greater success. This return could be attributed in part to the emergence of new training techniques that allows training deep networks efficiently. Obviously, progresses in hardware have also played an important role in this come-back: Powerful graphical processing units (GPU) are very useful when matrices and vectors operations are needed, leading to speed up of the training processes by orders of magnitude. Today, in many applications in signal processing and machine learning, neural networks are competing favorably with state of the art methods. In this work we have used neural networks for inverse problems in the context of image restoration. In the first part of this thesis, we harness this tool for image reconstruction in computed tomography. Our technique is based on a local nonlinear fusion of several image estimates, all obtained by applying a chosen reconstruction algorithm with different values of its control parameters. Usually, such output images have different bias/variance trade-off. Their fusion by a fully connected neural network is trained on a set of known examples, leading to a substantial improvement. Numerical experiments show an increase in reconstruction quality relatively to existing direct and iterative reconstruction methods, facilitating a n appealing alternative to commonly used methods for radiation reduction in CT imaging. In the second part of this thesis, we use neural networks to tackle the block-based compressed sensing problem. In compressed sensing the aim is to reconstruct a signal from a small number of measurements, obtained by linear projections of the signal. Block-based CS is a lightweight CS approach that is mostly suitable for processing very high-dimensional images and videos: it operates on local non-overlapping patches, employs a low-complexity reconstruction operator and requires significantly less memory to store the sensing matrix. In this work we present a deep learning approach for block-based CS, in which a fully-connected network performs both the block-based linear sensing and non-linear reconstruction stages. During the training phase, the sensing matrix and the non-linear reconstruction operator are jointly optimized, and the proposed approach outperforms state-of-the-art both in terms of reconstruction quality and computation time.
* MSc seminar under the supervision of Prof. Michael Elad & Prof. Michael Zibulevsky