2016-04-27 EECS 556 abstracts / schedule, 1-4pm, 1005 EECS #1 1:05 Spectral-spatial hyperspectral image segmentation Takeshi Kondoh, Anish Lahiri, Andrew Sender The paper we studied and reproduced is on a new supervised method for segmenting remotely sensed hyperspectral images which learns the spectral and spatial information in a Bayesian MAP framework. A multinomial logistic regression (MLR) algorithm is first used to learn the spectral information. The next step is learning the spatial information by multilevel logistic (MLL) Markov Random Field (MRF). The authors simulated one dataset and used two real world datasets to prove the success of their new method. We reproduced their results and additionally tried prefiltering the images before running the new method to find no significant changes. #2 1:25 Fast partitioning of vector-valued images and 3D volumes Jonathan Macoskey, Steven Parkison, Josiah Simeth This report explores a method of computationally efficient image segmentation. We explore an approach first presented by Storath et al. of splitting the Potts minimization problem using an approach based on Alternating Direction Method of Multipliers optimization. We show this approach out performs other state of the art approaches that are based on Graph-Cuts and Convex Relaxation. Finally we show two novel extensions of this method, one extends this approach into three dimensional medical imaging, and one shows how it can be easily parallelized for shorter run times. #3 1:45 Light field compression and reconstruction using sparsity in the continuous Fourier domain Nathan Sawicki, Mostafa Shuva Light field imaging provides a way to capture more information about a scene than traditional imaging. However, acquiring good light field datasets requires sufficiently sampling the light field, or taking images from sufficiently many perspectives. Once the light field is collected, the issue of storing the light field arises. We propose a compression and reconstruction scheme for reducing the storage size of a light field based on a similar reconstruction method proposed by Lixin Shi et al. at MIT. Our proposed method uses continuous Fourier domain analysis, sparsity, and optimization. #4 2:05 Local edge-preserving multiscale decomposition for high dynamic range image tone mapping Wonhui Kim, Hongki Lim A High Dynamic Range (HDR) image has a large ratio between the maximum and minimum intensities of the image. Since it usually exceeds the dynamic range of standard displays, tone mapping process is required. The key to HDR tone map- ping is to preserve details while compressing the unimportant image components. Therefore, most state-of-the-art approaches to HDR tone mapping involve sepa- rating an image into important and less important signal components. Motivated by the underlying principle of Human Visual System (HVS), which states that human vision is more sensitive to high-frequency components of the image rather than the low-frequency components, it is common to consider low frequency components less critical where minor distortion does not affect the qual- ity of final tone-mapped images. Based on such idea, the paper presents an HDR tone mapping framework that first separates an image into a piecewise smooth base layer and detail layers using a novel Local Edge Preserving (LEP) filter. [break for 5 minutes] #5 2:30 A general framework for patch-based image denoising Ravi Chandra Donapati, Enpei Xi, Yafang Xue The search for an efficient image denoising method remains a valid challenge. To address this problem, we replicate the work from "A tour of modern image filtering". By making contacts with the field of non-parametric statistics, Milanfar etc. proposed a general framework that combines several existing filtering algorithms that enables performance analysis and subsequent improvement of these algorithms. Specifically, we implemented classical kernel regression, bilateral filtering, non-local means and locally adaptive kernel regression (LARK) for denoising. A subsequent diffusion based iteration method was used to improve the performance of these algorithms. We also proposed a fast non-local means algorithm to accelerate the NLM algorithm using K-means clustering, where the weights are calculated from clustered patches. Additionally, we explored the application of kernel regression on image deblurring and discussed the results. #6 2:50 An iterative shrinkage approach to total-variation image restoration Zhenzhen Dai, Ahmet Mazacioglu, Junyan Shi Image restoration is a vital image processing form that aims to recover clean, true images from noisy, blurry, corrupt images. This is an ill-posed problem that requires a priori information about the original image. A common case is when the original image is known/assumed to be piecewise smooth, then a certain set of standard priors lead to the highly popular total variation (TV) based image restoration techniques. In this project, we have picked a TV- based image restoration technique that incorporates a new iterative shrinkage method (TVIS) which employs recursive linear filtering and soft thresholding. Restoration performance of TVIS is measured over a range of images, blur types, and noise levels. Finally, TVIS is quantitatively compared to another TV-based restoration algorithm in literature that is taken as our reference model. #7 3:10 Single Image blind image deconvolution Yifan Ji, Tairan Liu, Ningyuan Wang Image blurring caused by camera shake can be categorized as an image deblurring problem. Our project attempted to reproduce a blind deconvolution method proposed by Shan. The method updates deblurred image and kernel iteratively, based on an author-defined probability model. We attempted to implement their work, and made one adjustment for finding f by observing f should be non-negative. The result is mixed, failing to reproduce their result for the complete blind deconvolution problem, but parts of the code are tested to give partial results. [break for 5-10 minutes, then project award(s)]