Anisotropic edge-enhanced active contour model with Gaussian difference for robust multi-category image segmentation
摘要
Active contour model (ACM) is a pivotal approach for image segmentation. However, conventional models are susceptible to noise and weak target boundaries, limiting their practical applications. Furthermore, uncertainty in the initial contour heightens model sensitivity. To address these challenges, we introduce DoG&EED, an active contour model grounded in the anisotropic diffusion equation and Gaussian convolution. The proposed anisotropic diffusion equation, termed EED, enhances noise robustness and edge localization by strengthening gradients along target boundaries while weakening them in noisy regions. Leveraging Gaussian kernel functions, we construct a streamlined energy function, the DoG operator, which efficiently captures essential image information while mitigating model complexity. A novel regularization function, independent of the global energy function, preserves its effectiveness while reducing iteration costs. By integrating YOLOv5 bounding boxes with DoG&EED, we mitigate initial contour sensitivity. Extensive experiments validate that DoG&EED boasts robust anti-noise capabilities and maintains high segmentation accuracy, achieving 81.2 DSI, 76.9 SI, and 78.4 JI in homogeneous experiments, along with 79.3 mIoU, 46.5