Deep Image Segmentation via Discriminant Feature Learning
This work addresses the limitations of existing image segmentation methods, which often produce blurry boundaries and low-confidence predictions due to loss functions that neglect the discriminative structure of features. To overcome this, we propose a differentiable, architecture-agnostic Deep Discriminant Analysis (DDA) loss that integrates classical discriminant analysis principles into end-to-end training. The DDA loss explicitly maximizes inter-class variance while minimizing intra-class variance, thereby encouraging the learning of compact and well-separated feature representations. Notably, it incurs no additional inference overhead and can be seamlessly incorporated into any segmentation network. Extensive experiments on the DIS5K benchmark demonstrate that DDA consistently enhances segmentation accuracy, boundary sharpness, and model confidence, confirming the effectiveness of discriminative feature learning for image segmentation.