Unsupervised, Untrained, and Robust Single Image Superpixel Segmentation Network

Image superpixel segmentation has greatly benefited from the excellent feature extraction capabilities of neural networks. However, most existing neural network-based superpixel segmentation methods require large amounts of labeled data for training, which limits their generalization and practical applications. To address this issue, this paper introduces a novel unsupervised network named UURS for robustly generating accurate superpixels from a single image without requiring an image collection (untrained) and any labeled data (unsupervised). First, an efficient feature extraction module based on Lipschitz-controlled convolutional layers and a channel attention mechanism is proposed to leverage deep image prior for capturing discriminative features. Furthermore, an image reconstruction loss combining structure- and pixel-level similarity is designed to improve feature accuracy. Subsequently, a classification module generates superpixels from these features. The network achieves optimal performance through an adaptive stopping criterion that prevents underfitting or redundant computation. We validate UURS through extensive experiments on standard image segmentation benchmarks. Additionally, synthetic experiments with diverse noise distributions in sRGB space and evaluations on real-world noisy image datasets demonstrate the network’s robustness. Results show that our method outperforms state-of-the-art approaches in boundary adherence and maintains robustness across both clean and noisy conditions.