Threshold Image Segmentation Algorithm Based On Adaptive Histogram Statistical Model

In most of computer vision applications image segmentation has been a fundamental and primary step with a great impact on the analysis and processing of the next step. At the same time, it is also a classic dilemma in image processing applications. Many techniques are available for image segmentation and histogram thresholding method is one of the most widely used techniques. This paper bases on the traditional multi-threshold segmentation algorithm using Gaussian fitting, which has some defects, combining the probability and mathematical statistics to propose two other image segmentation algorithms:adaptive multi-threshold segmentation algorithm with Gamma distribution and adaptive multi-threshold segmentation algorithm using ISODATA with Log-Normal distribution.The multi-threshold segmentation algorithm based on Gaussian fitting uses Gaussian distribution to approximate the histogram, but the skewness of the Gaussian distribution is zero and a symmetric distribution, which can only model the symmetric modes. When the skewness of the samples isn't zero and the histogram of an image is non-symmetric, segmentation of Gaussian fitting method will be imprecise. Consequently, this paper will propose Gamma distribution, which can model symmetric and non-symmetric modes, to approximate the histogram, and combine the ISODATA technique to adaptively decide the threshold values. The algorithm firstly uses the Gamma distribution to estimate the parameters required to apply the maximum likelihood estimator, so it avoids the interference caused by multi-peak overlapping; then it splits the histogram (non-homogenous region) into distinct homogenous regions and merges the adjacent classes if they apply the merge constrains, which is called histogram split-merge technique. The technique avoids the drawbacks of normal split-merge technique like compact region and the rough edge when applied directly to the image; at last the method yields better segmented images with non-symmetric histogram. Experimental results showed that:when the histogram is almost symmetric the proposed method, which based on Gamma distribution and the existing method, which is based on Gaussian distribution gives almost the same results; however, when the histogram is non-symmetric, Gamma distribution gives better result than Gaussian distribution, the fitting precision of the histogram and the robustness of image segmentation have both been effectively improved.Gaussian distribution can model only the symmetric modes, and Gamma distribution can model symmetric and non-symmetric modes. However, some distributions, whose skewness isn't zero, especially is positively skewed, the paper proposes an adaptive multi-threshold segmentation algorithm using ISODATA with log-normal distribution. The method firstly uses the Log-Normal distribution to estimate the parameters required to apply the moment estimator. Then it using homogeneous tests and merge steps with histogram splits and merges classes. At last the proposed method is tested at parameter estimation and images, which are added log normal noise, whose skewness is positive. The experimental results show, the log normal distribution can model symmetric and positively skewed (non-symmetric) data, so the proposed method which based on log normal distribution yields better segmented image than existing methods which rely on Gaussian and Gamma distributions.