Study On Region Based Image Segmentation Technique

Image segmentation is a key problem in computer vision. In this paper we mainly studied the region-based image segmentation technique. In methodology we make our research in two aspects: one is the transition region extraction (TRE) based image segmentation technique, the other is the graph theory based image segmentation technique. Most of our works focus on the former technique.By analyzing the drawbacks of the traditional transition region extraction methods we give a more generalized definition on transition region, and suggest that the searching aspects of TRE should be converted to study methods that can make transition region directly extracted. Besides we give clear explanation of the thresholding criteria geometrically and mathematically.In gradient based TRE method study we present degree based TRE and image segmentation method (D-TREM), and wavelet energy ratio based TRE and segmentation method (W-TREM). In D-TREM we consider an image as an undirect graph and extract transition region by pixel's degree information. The scale parameter of D-TREM make it has certain ability to deal with noise. W-TREM is only better for segmenting special type of texture images than for other kind of images. Its computation speed is slow. D-TREM is better than other gradient-based TRE method. In non-gradient TRE method study we present local entropy based TRE and image segmentation method (LE-TREM), and local complexity based TRE and image segmentation method (C-TREM). Both LE-TREM and C-TREM have good performance to deal with salt and pepper noise, which ability comes from their way measuring grayscale level information. C-TREM is faster than LE-TREM, and has no small sample problem.When there are several different kinds of objects in an image, the histogram of transition region will be multi-modal distribution. Each mode of the histogram will be corresponding to a possible threshold. We adopt a non-maxima suppression filter to detect the location of the mode. A criterion considering both discrepancy and bites is adopted for determining optimal classification numbers and multilevel thresholds.Segmentation evaluation is performed completely on four TRE methods and two traditional thresholding methods. We make our evaluation on several aspects including segmentation quality, ability to deal with noise, computational complexity, object/background ratio and number of manual parameters in certain algorithm. Evaluation results demonstrate that C-TREM is a promising method in all TRE methods and other two traditional methods. It can help to segment wide range of images. The drawbacks of traditional TRE methods are also proved in the evaluation work.A novel graph theory based histogram clustering thresholding (GHCT) method is proposed. Traditional graph theory based image segmentation algorithm proceeds directly on 2D image data, which is very slow in computation time. GHCT make clustering operate on data of histogram. The computing expense is limited on 256 histogram data and has no relationship with the image size.