A comparison of the performances of a bayesian algorithm and a kohonen map for clustering texture data
暂无分享,去创建一个
With many clustering algorithms available, it may be difficult to discern which is better for a given task. This study compares the performance of two clustering algorithms, the Bayesian classifier AutoClass and a Kohonen map, for the task of identifying classes of different textures in images based on statistics derived from gray-level co-occurrence matrices. The performance of the two algorithms is assessed in terms of quality of the classification. Comparisons of quality are given in terms of objective criteria such as cluster diameter, intercluster distance, etc. as well as subjective judgements by domain experts. Two different types of images are used. The first type of image consists of standard texture images in which textures classes are readily identified by novices. The second type consists of side-scanned sonar images in which the clusters are not necessarily apparent to novices and are not always classified consistently by domain experts (geologists). INTRODUCTION With many clustering algorithms available, it may be difficult to discern which is better for a given task. This study compares the performance of two clustering algorithms, the Bayesian classifier AutoClass and a Kohonen map, for the task of identifying classes of different textures in images based on statistics derived from gray-level co-occurrence matrices. The performance of the two algorithms is assessed in terms of quality of the classification. Comparisons of quality are given in terms of objective criteria such as cluster diameter, intercluster distance, etc. as well as subjective judgements by domain experts. Two different types of images are used. The first type of image consists of standard texture images in which texture classes are readily identified by novices. The second type consists of side-scanned sonar images in which the clusters are not necessarily apparent to novices and are not always classified consistently by domain experts (geologists). Many different techniques have been developed to classify textures in images. Since texture is a characteristic of groups of pixels rather than individual pixels, any method that identifies texture classes in images must either subdivide the image into homogeneous regions and then classify these regions or must classify each pixel based on characteristics of its neighbors. In our research, our goal is to “province” images of the ocean floor based on visual
[1] T. Reed,et al. Digital image processing techniques for enhancement and classification of SeaMARC II side scan sonar imagery , 1989 .
[2] M.,et al. Statistical and Structural Approaches to Texture , 2022 .
[3] Kishan G. Mehrotra,et al. Elements of artificial neural networks , 1996 .
[4] Peter C. Cheeseman,et al. Bayesian Classification (AutoClass): Theory and Results , 1996, Advances in Knowledge Discovery and Data Mining.