PENERAPAN TRANSFORMASI WAVELET UNTUK REDUKSI SPEKEL, ESTRAKSI CIRI DAN SEGMENTASI CITRA BERDASARKAN TEKSTUR

• Main Author: Kun Wardana Abyoto, R.
• Format: Dissertations
• Language: Indonesia
• Subjects: Fisika
• Online Access: https://digilib.itb.ac.id/gdl/view/5537
• Institution: Institut Teknologi Bandung

<b>Absrack</b><p align=\"justify\">In this dissertation, we propose the implementation of wavelet transform in image processing field that covers low level, middle level and high level processing, that is, speckle reduction, feature extraction and texture segmentation, respectively. The wavelet transform has many properties that are beneficial to these three types of application, such as multiresolution representation, localized frequency information, direction sensitive of edge information, orthogonality, efficient algorithm and less of memory requirement. <br /> <br /> In the first application, namely speckle reduction for SAR imagery, many techniques have been employed to reduce speckle, but the qualifying factors that must be achieved by speckle reduction filter, such as suppression of speckle in a uniform area and preservation of edges, can hardly be satisfied simultaneously. When we add visually natural appearance as the third requirement, speckle reduction filters well balanced in these three criteria are a few. The implementation of multiresolution analysis of wavelet transform in speckle reduction process is intended to satisfy each of these criteria. The detail images produced by two-dimensional wavelet transform will be utilized to identify edge information in each scale. The proposed filter intends to suppress speckle by reducing the amplitude of the detail images in wavelet subspaces, while preserving edges by releasing the amplitude reduction around edges. By reconstructing the modified detail images using inverse wavelet transform, the filter will result in a smoothed image without blurring edges. Simulations and implementation to SAR images have indicated that the speckle from original image can be suppressed significantly without changing its mean value and without losing the details and narrow edges. Comparing this performance with the result of the highly publicized filtering algorithms such as median filter, Lee filter and weighted filter, shown that the proposed algorithm is more superior in both smoothing and edge preservation, and in generating visually natural images as well. <br /> In the second application, namely textural feature extraction, many traditional feature extraction methods are primarily focus on the coupling between image pixels on a single scale, so that they are not sufficient to characterize different scales of textures effectively. The other main difficulties that frequently arises are the redundancy of information represented in the features, and the number of dimensionality which is usually quite large, leading to a degradation in the performance of the classifiers. The implementation of wavelet transform in textural feature extraction process aims to optimize the textural discriminatory information, whilst, at the same time, performing in a small dimension of feature space. The proposed feature extraction method leads to the concept of muitiresolution analysis, so that we can analyze the signal at successive scales or resolutions. Furthermore, we are able to zoom into any dominant frequency adaptively. This method consists of two main steps: wavelet transforms and texture measure computation process. In this proposed method, the tree-structured wavelet transform is optimized by assigning the channels that have maximum texture information at each resolution level as texture features, so that only significant texture elements is extracted from the given image. The implementation of this method in supervised image classification for 25 natural texture classes yields an overall accuracy rate of 99.12%. In addition, this method provides quick response that takes only 0.4 second in accomplishing the feature extraction process, for each texture samples of size 128 x 128 pixels. In its entirety, this experimental result indicated that the proposed method could give a reasonable result, while improving the computation performance in time. <br /> <br /> The third application, namely unsupervised texture segmentation, is an extension of the proposed textural feature extraction method. This texture segmentation procedure consists of two steps: feature extraction and clustering process. The process in the first phase results in a set of feature images that contains a set of feature vectors. These feature vectors corresponding to the different resolution of decomposed images are assumed to capture and characterize effectively different scales of textures from the input image. In the second phase, all the extracted features are classified based on their associated vector values by using fuzzy K-means clustering algorithm. This clustering algorithm is useful especially in handling the ambiguity or uncertainty around the boundaries between two or more neighboring textures. This second phase will result in a segmented image whose regions are distinct from one another with respect to the texture characteristic content. The main advantage of this proposed procedure is that these computations can be performed in a lower-dimensional space that essentially preserves the discriminative information and provides features that are approximately decorrelated. This will simplify the segmentation process, while not seriously affecting the overall performance. From the simulation using synthetic image, the proposed procedure could achieve the segmentation result with an error rate of 3.21%. In addition, the implementation of the procedure using SAR image give satisfying result in discriminating the regions that have different texture characteristics. <br /> <br /> In this study, we also examine the discrimination ability of four different wavelet bases: Haar, 4-tap Daubechies, 16-tap Daubechies, Battle Lemarie, and three different texture measures: energy, uniformity and entropy. From the experiment, it is shown that 16-tap Daubechies wavelet basis and energy measure give the best performance for textural feature extraction both in supervised image classification and unsupervised image segmentation. <br /> <br /> From all of the three proposed applications, it is shown that the approach using \'wavelet transform has performed quite satisfactory. Observing from the result, it may be reasonable to replace the traditional techniques in speckle reduction, textural feature extraction and texture segmentation application that still utilizing single resolution analysis methods or still using Fourier transform, in order to benefit by a better compromise between quality of the processing result, computation time and computer memory requirement.