Multilevel Thresholding for Image Segmentation based on Similarity Filtering

Objective: In this paper, an effective and fast multi-threshold image segmentation method is proposed based on similarity filtering. Method: The image histogram peaks and the valley can be used to locate the clusters in the image. The idea of the proposed research is to fit the Gaussian distribution to the histogram of the image. Dominant peaks are selected from the input image histogram near to its Gaussian distribution. Then for each element of the peaks, peak’s valleys are obtained in the left (low) and right (high) side. Findings: Experiments on a variety of images from Berkeley Segmentation Dataset (BSD) show that the new algorithm effectively segments the image in a computationally efficient manner. Comparison/ Performance evaluation: On comparison, proposed approach is found to be better than other existing methods. Peak Signal to Noise Ratio (PSNR) and time are used to evaluate the performance. The proposed algorithm tries to fit Gaussian curves on the dominant peaks and thus find the valleys which are used as thresholds. Novelty: This is always a quicker process as there is a predefined model which only needs to be fit for the given data set.