PA3: Contributed Session

Room: Old Main Academic Center 3070

Webex Link

Chair: Hyeona Lim

Chartese Jones, University of Missouri

Time: 4:00 pm - 4:20 pm (CST)

Title: Quarter Match Non-Local means Algorithm for Noise Removal


Image denoising is an essential, nonetheless stimulating dilemma that has been investigated for decades, and there always remains room for improvement. Understanding how noise can be present in images have led to multiple denoising techniques. Some research shows the non-local means  ltering method outperform many denoising algorithms/techniques. The non-local means  lter removes noise and sharpen edges without losing too many  ne structures and details. Also, the non-local means algorithm takes advantage of the high degree of redundancy of any natural image and is very accurate since all pixels contribute for denoising at any given pixel. However, a major drawback for the non-local means  lters is the computational burden. This is due to the non-local averaging. We investigated innovative ways to reduce the computational burden. We propose non-local means based denoising  lter for images contaminated by noise, which a ects multiple pixels in images, to aid in this task. When denoising images there is a battle between noise reduction and preservation of actual features, which makes reduction of noise a difficult task. Because of this understanding, we propose a quarter match non-local denoising algorithm that is very e ective in noise removal. In this method, we selectively calculate the weights of the non-local means filters by dividing each patch into four quarters, where the filters help to classify a more concentrated region in the image patches, and thereby reduce the computational burden and producing a superior comparison for overlying.

(joint work with Hyeona Lim)

Ben Spencer, Indiana University

Time: 4:20 pm - 4:40 pm (CST)

Title: Improvements of Sapiro's Selective Non-Local Means filter method


Images occur in many forms in our daily lives, and denoising is one of the most important steps of image analysis in improving the quality of images. It is also used as a preprocessing for other imaging techniques such as feature segmentation and compression. One of the more recent techniques among the various image denoising methods developed by Buades, Coll, and Morel in 2005 is the Non-Local Means (NLM) filter. In this method, each pixel is denoised using a weighted average of all pixels in the image. This method becomes quite useful when dealing with images with repetitive patterns or fine structure. However, the method is computationally expensive due to the non-local averaging. To reduce the computational burden, Sapiro improved the NLM method by adding in a selective process where the weighted average of the pixels who were not "similar" enough to the original pixel was discarded in the computation. In this talk, I will discuss the NLM method, Sapiro's improvement on the NLM method, as well as the new improvements that allow some of the constant parameters to vary as the pixel's "similarness" varies.

Shiraz Mujahid, Mississippi State University 

Time: 4:40 pm - 5:00 pm (CST)

Title: Parametric study of variational and non-local image denoising methods


Given the ubiquity of image analysis and rapid post-processing in commercial as well as research applications, the ability to denoise images effectively is of key importance. This study investigates the effects of controlling parameters for two broad categories of denoising methods. The first utilizes a calculus-of-variations based approach, treating the desired clean image signal as an optimal functional, with the noise assumed to be a Gaussian additive perturbation superimposed over the noisy image functional. Implementation parameters such as finite difference scheme, number of iterations, and amount of regularization can be varied to modulate the convexity of the input functional as well as the extent of local diffusive denoising. The second series of methods involve non-local means filtering, which relies upon reinforcement of self-similar features within the input image to generate a denoised image while reducing non-self-similar spurious noise. Here, various selection criteria are introduced to quantify self-similarity, and the influences of these selective parameters on the effectiveness of the denoising implementation are established. The effectiveness of each denoising implementation is measured using three metrics: computational time, peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM). The characteristics of PSNR and SSIM as measures for quantifying image denoising are compared, and the parameters used within the denoising methods are varied to provide optimal balance between PSNR and SSIM for multiple categories of noisy images.

(joint work with Hyeona Lim)