University of Tübingen, 2015. — 142 p.Image blur is a fundamental problem in both photography and scientific imaging. Even the most well-engineered optics are imperfect, and finite exposure times cause motion blur. To reconstruct the original sharp image, the field of image deconvolution tries to recover recorded photographs algorithmically. When the blur is known, this problem is called non-blind deconvolution. When the blur is unknown and has to be inferred from the observed image, it is called blind deconvolution. The key to reconstructing information lost due to blur and noise is to use prior knowledge. To this end, this thesis develops approaches inspired by machine learning that include more available information and advance the current state of the art for both non-blind and blind image deconvolution. Optical aberrations of a lens are encoded in an initial calibration step as a spatially-varying point spread function. With prior information about the distribution of gradients in natural images, the original image is reconstructed in a maximum a posteriori (MAP) estimation, with results comparing favorably to previous methods. By including the camera’s color filter array in the forward model, the estimation procedure can perform demosaicing and deconvolution jointly and thereby surpass the quality of the results yielded by a separate demosaicing step. The applicability of removing optical aberrations is broadened further by estimating the point spread function from the image itself. We extend an existing MAP-based blind deconvolution approach to the first algorithm that is able to remove spatially-varying lens blur blindly, including chromatic aberrations. The properties of lenses restrict the class of possible point spread functions and reduce the space of parameters to be inferred, enabling results on par with the best non-blind approaches for the lenses tested in our experiments. To capture more information about the distribution of natural images and capitalize on the abundance of training data, neural networks prove to be a useful tool. As other successful non-blind deconvolution methods, a regularized inversion of the blur is performed in the Fourier domain as an initial step. Next, a large neural network learns the mapping from the preprocessed image back to the uncorrupted original. The trained network surpasses results of state-of-the-art algorithms on both artificial and real-world examples. For the first time, a learning approach also succeeds in blind image deconvolution. A deep neural network “unrolls” the estimation procedure of existing methods for this task. After training end-to-end on artificially generated example images, the network achieves performance competitive with state-of-the-art methods in the generic case, and even goes beyond when trained for a specific image category.
Чтобы скачать этот файл зарегистрируйтесь и/или войдите на сайт используя форму сверху.