Otsu Thresholding Model Using Heterogeneous Mean Filters for Precise Images Segmentation

Digital image segmentation can be performed using different approaches, such as machine learning, classification, or low-level image processing. Otsu’s method is a frequently used technique for histogram thresholding-based image segmentation under a low-level image processing approach. Otsu’s algorithm finds the threshold value by maximizing the objective function and this process relies on the sum of normal distribution for intensities in the image histogram. There are different forms of images with various structures of intensity distribution that may not fit with Gaussian-based Otsu. This paper aims to estimate enhanced mean values for the Otsu algorithm and improve it to be compatible with various types of images for better segmentation output. Medical Resonance Imaging (MRI) brain tumor and Dermoscopic skin lesion images have been used for segmentation. The proposed model uses existing mean filter approaches heterogeneously. In other words, selecting and combining two different mean filters to estimate the mean value, as before each image region has its own mean. The main aim is to handle the poor quality of the images when estimating the mean value for Otsu’s between-class variance. The proposed method has been tested beside the original Otsu method and literature-related works. The proposed model showed improved results based on unsupervised and supervised evaluation of segmentation results.