Amirkeyvan Momtaz

PhD Thesis Defense Session

Abstract:
In non-destructive testing, detection and clustering of defects is an important issue. One of the exploited methods to determine the defects is the use of c-scan images obtained from ultrasound test. The goal of the thesis is detection and clustering of defetcts in ultrasound images. Since the quality of the obtained image is not suitable for processing, it is necessary to enhance the quality of images before applying clustering method. The proposed denoising method in preprocessing step is based on the denoising the wavelet coefficients of the image by the use of independent component analysis and a spatial filter. The filter is used to determine the homogenous areas from the areas containing image details. The method has the capability to reduce different kind of noises including Gaussian, speckle and the noise with weak Gaussian distribution.
The proposed clustering algorithm is based on the rosette pattern. For this purpose, by the use of the rosette pattern, the image is sampled and according to the rosette pattern characteristics, the samples are mapped to the two dimensional linear space. In this stage, based on the neighborhood property of the samples, the clustering is performed. Finally, the clustered samples are remapped to the main space. Unlike the conventional clustering methods such as k-means and FCM algorithms requiring the number of clusters as one of the initializing parameters, in the proposed method, there is no need to initialize any parameter. Based on different data sets, the results show that the algorithm improves the capability of clustering, run time and determining the optimal number of clusters about 92%, 99% and 71% compared to k-means and FCM algorithms, rspectively. Moreover, in dealing with high resolution data sets, the efficiency of the algorithm in clusters detection and run time improvement increases considerably.

Day: Wednesday, Date: 2012/01/11 Time: 17
Class: 303