Image Segmentation Enhancement using aggregated Kernel Graph Cut

Abstract

Kernel graph cut-based image segmentation methods have demonstrated effectiveness in processing image data, yet they often encounter limitations when handling complex image structures due to inherent constraints. The conventional kernel graph cut algorithm is limited in effectively handling intricate image structures. These limitations arise from its inability to adequately capture complex image features. This paper addresses these limitations by proposing an innovative solution that aggregates spatial and intensity information within a kernel framework. The methodology introduces a new construction of the kernel based on the fusion of spatial and intensity data, generating maps that delineate differences in both pixel intensity and spatial relationships. These maps contribute to generating a comprehensive similarity matrix. Leveraging the RBF kernel, data mapping is improved, enhancing the efficacy of the similarity matrix. Through this aggregated approach, the constraints of traditional kernel graph cut methods are surmounted. This enhancement significantly boosts the Dice coefficient and accuracy of image segmentation, particularly in scenarios featuring complex image structures and diverse intensity distributions. The proposed methodology exhibits promising results, showcasing a 99.40% Dice similarity, surpassing standard energy-based methods.