Image Segmentation Based on an Improved Fuzzy Clustering Algorithm

Abstract

Traditional fuzzy clustering algorithms are considered powerful tools for image segmentation. However, these algorithms face two main challenges. First, they are sensitive to outliers. The fuzzy memberships in these algorithms are non-dispersive, meaning they are heavily influenced by outliers, largely due to the use of squared error in their objective function. This flaw can lead to incorrect and unreliable clustering results, reducing robustness. Second, they tend to produce an excessive number of clusters. Traditional fuzzy clustering algorithms often create too many clusters, many of which are unnecessary and redundant. This phenomenon, known as over-segmentation in fuzzy clustering, occurs due to the image's loss of local spatial information. To address these challenges, this study presents a solution that enhances the robustness of the fuzzy clustering algorithm. The proposed algorithm includes two main components: the first involves adding a Gaussian-based regularizer to the objective function, which incorporates a Gaussian sub-criterion to calculate the distance between data points and cluster centres. By adding this criterion, the proposed method increases the dispersion of fuzzy membership functions, thereby reducing the impact of outliers and improving clustering accuracy. The second component involves using a filter to resolve the problem of excessive clustering. The proposed algorithm was compared with traditional fuzzy clustering methods and spatial information-based methods to validate its performance, yielding superior results. The algorithm achieves higher accuracy and cohesion in image segmentation while being more robust to outliers and noise.