Variational segmentation model with non-local image dehazing for boundary extraction in vector-valued hazy images

Abstract

Image segmentation is a fundamental technique for delineating boundaries within images, enabling detailed analysis and simplifying subsequent processing tasks. Recently, a variational selective segmentation model for vector-valued (color) images, termed Selective Segmentation based on Gaussian Regularization (SSGR), was introduced. In this variational model, the standard regularization term is replaced with a Gaussian function, yielding a speed improvement of approximately 247 times compared with the earlier formulation, while maintaining comparable accuracy. However, the SSGR model shows limitations when applied to hazy images. To address this, the present study reformulates the SSGR model by incorporating Non-Local Image Dehazing as the new fitting term, resulting in a modified approach named Selective Segmentation for Haze Image (SSHI). Experiments were conducted on hazy images to evaluate the proposed model. The performance of SSHI was assessed using common segmentation metrics: Error, Accuracy (ACU), Jaccard Similarity Coefficient (JSC), and Dice Similarity Coefficient (DSC). Efficiency was measured by recording processing time. Numerical experiments demonstrated that the SSHI model, through the integration of Non-Local Image Dehazing, consistently achieved the highest ACU, JSC, and DSC values, while also yielding the lowest Error and processing time compared to existing methods. These findings confirm that SSHI provides improved accuracy and efficiency for hazy image segmentation. The proposed model has potential applications in domains where hazy imaging is common, including medical imaging such as X-ray and endoscopy, remote sensing, and environmental surveillance. Future research may extend this framework to three-dimensional formulations for more complex imaging applications.