Jyothisha J. Nair

Denoising of SAR images using Maximum Likelihood Estimation

Image denoising is an important problem in image processing because noise may interfere with visual interpretation. This may create problems in certain applications like classification problem, pattern matching, etc. This paper presents a new approach for image denoising in the case of speckle noise models. The proposed method is a modification of Non Local Means filter method using Maximum Likelihood Estimation. The Non Local Means algorithm performs a weighted average of the similar pixels. Here we introduce a method that performs weighted average on restricted local neighborhoods. More over the method performs weight calculation using Geman-McClure estimation function rather than the exponential function because of the fact that Geman-McClure estimator is better in preserving edge details than the exponential function. Experiments at various noise levels based on PSNR values and SSIM values show that the proposed method outperforms the existing methods and thereby increasing the accuracy of further processing for synthetic aperture radar (SAR) images.

Improvised Apriori with frequent subgraph tree for extracting frequent subgraphs

Graphs are considered to be one of the best studied data structures in discrete mathematics and computer science. Hence, data mining on graphs has become quite popular in the past few years. The problem of finding frequent itemsets in conventional data mining on transactional databases, thus transformed to the discovery of subgraphs that frequently occur in the graph dataset containing either single graph or multiple graphs. Most of the existing algorithms in the field of frequent subgraph discovery adopts an Apriori approach based on generation of candidate set and test approach. The problem with this approach is the costlier candidate set generation, particularly when there exist more number of large subgraphs. The research goals in frequent subgraph discovery are to evolve (i) mechanisms that can effectively generate candidate subgraphs excluding duplicates and (ii) mechanisms that find best processing techniques that generate only necessary candidate subgraphs in order to discover the useful and desired frequent subgraphs. In this paper, a two phase approach is proposed by integrating Apriori algorithm on graphs to frequent subgraph (FS) tree to discover frequent subgraphs in graph datasets.

A robust non local means maximum likelihood estimation method for Rician noise reduction in MR images

Denoising is one of the most important preprocessing task in medical image analysis. It has a great role in the clinical diagnosis and computerized analysis. When SNR is low, medical images follows a Rician noise distribution which is signal dependent. In the literature, only few works focus on the edge preserving quality of MR images. Our aim is to estimate the noise free signal from MR magnitude images by focusing on preserving edges and tissue boundaries. The proposed method is an improvisation over non local means maximum likelihood approach for Rician noise reduction in MR images. Our method focus on a robust estimator function (Geman-McClure function) for weight calculation, and is compared with the existing methods in terms of PSNR ratio, visual quality comparison and by SSIM values. The proposed method outperforms the state-of-the art methods in preserving fine structural details and edge boundaries.

Speckle Noise Reduction Using Fourth Order Complex Diffusion Based Homomorphic Filter

Filtering out speckle noise is essential in many imaging applications. Speckle noise creates a grainy appearance that leads to the masking of diagnostically significant image features and consequent reduction in the accuracy of segmentation and pattern recognition algorithms. For low contrast images, speckle noise is multiplicative in nature. The approach suggested in this paper makes use of fourth order complex diffusion technique to perform homomorphic filtering for speckle noise reduction. Both quantitative and qualitative evaluation is carried out for different noise variances and found that the proposed approach out performs the existing methods in terms of root means square error (RMSE) value and peak signal to noise ratio (PSNR).

Image convolution optimization using sparse matrix vector multiplication technique

Image convolution is an integral operator in the field of digital image processing. For any operation to be processed in images say whether it is edge detection, image smoothing, image blurring, etc. process of convolution comes into picture. Generally in image processing the convolution is done by using a mask known as the kernel. As the values of the kernel is changed the operation on image also changes. For each operation, the kernel will be different. In the conventional way of image convolution, the number of multiplications are very high. Thereby the time complexity is also high. In this paper, a new and efficient method is proposed to do convolution on the images with lesser time complexity. We exploit the sub matrix structure of the kernel matrix and systematically assign the values to a new H matrix. Since the produced H matrix is a spare matrix, the output is realized here by using Sparse Matrix Vector Multiplication technique. Compressed Row Storage format (CSR) is the format that is used here for the Sparse Matrix Vector Multiplication (SMVM) technique. Using the CSR format with Sparse Matrix Vector Multiplication technique, convolution processes achieves 3.4 times and 2.4 times faster than conventional methods for image smoothing and edge detection operations respectively.

All pair shortest path using distributed architectures

Due to the explosion of data from various sources, data analytics is found to be difficult using the CPU alone. For huge networks, the most popular graph algorithms using a single processor failed to accomplish this task. Hence the need of algorithms that have higher processing capabilities became dominant. Graph analytics is gaining importance in the realm of data analysis due to the advantages over other conventional analytical methods. All pair shortest path problem or path-finding problem has applications in various fields. Finding the all pair shortest path in a graph is computationally complex and time consuming in the case of large networks. In this paper, we propose a map reduce approach using the in-memory computation for finding the all pair shortest path using the transitive closure property and the greedy technique in Dijkstras single source shortest path method. In order to overcome the scalability issues in the network representation, we use a tuple based network representation. The performance of the proposed method is proved experimentally.

A survey on extracting frequent subgraphs

Mining on graphs has become quiet popular because of the increasing use of graphs in real world applications. Considering the importance of graph applications, the problem of finding frequent itemsets on transactional databases can be transformed to the mining of frequent subgraphs present in a single or set of graphs. The objective of frequent subgraph mining is to extract interesting and meaningful subgraphs which have occurred frequently. The research goals in the discovery of frequent subgraphs are (i) mechanisms that can effectively generate candidate subgraphs excluding duplicates and (ii) mechanisms that find best processing techniques that generate only necessary candidate subgraphs in order to discover the useful and desired frequent subgraphs. In this paper, our prime focus is to give an overview about the state of the art methods in the area of frequent subgraph mining.

Complex difffusion based image inpainting

Image inpainting is the meaningful reconstruction of lost, damaged or unwanted portions of an image by using the information from the proper undamaged portions of the same image. Image inpainting is an important process in image pro- cessing and has numerous applications in heritage conservation, restoration of old photographs, removal of occlusions, special effects in photos and so on. Here we replace the unwanted object by the information available from its neighbourhood. We present an algorithm that improves both the clarity and speed using a Complex-diffusion based approach. Complex-diffusion based approach for inpainting overcomes the shortcomings such as stair- case effect and excessive blurring caused by Partial Differential Equation based approaches. Our method outperforms the existing methods in terms of PSNR, SSIM and UIQI values.

Robust Non-Local Total Variation Image Inpainting

Image in painting is the process of removing selected object and restoring dead pixel from an image based on the background information. Various method have been proposed to tackle the in painting problem where they need related information from other images and use only neighboring data to recover the lost part of image. To overcome this, an efficient in painting technique called Robust Non-Local Total variation Method (RNLTV) is used. For filling lost portion, the proposed method uses information from the image itself and also superiority of the local and non-local methods are put together here. The local method which is efficient for recovering image edges and the textured region is recovered using nonlocal method. A Bregman operator splitting algorithm is employed here to avoid the loss of signal in each iteration of the total variation. The efficiency of the Robust Non-Local Total Variation method was tested and compared with existing methods and found superior.