Differentiable double clustering with edge-aware superpixel fitting for unsupervised image segmentation

Abstract

Unsupervised image segmentation is an essential topic in the field of computer vision, which broke the limitation of training data and expanded application scenarios. Off-the-shelf clustering methods simply rely on semantic concepts and incomplete boundary cues, resulting in incorrect segmentation in object boundaries. Therefore, this paper proposes an unsupervised image segmentation framework combining differentiable double clustering (DDC) and edge-aware superpixel (EA), which outperform prior work on the accuracy of the prior art. First, a multi-layer feature extraction network is introduced to combine low-level and high-level features for clustering. Then, the DDC module is designed to obtain initial labels of pixels from both local and global perspectives to improve the clustering accuracy. Pixel-wise feature similarity in different classes is pushed away, and one in the same class is brought closer. Finally, we use EA to provide well-fitting boundary cues for DDC label fusion to reduce incorrect segmentation. Extensive experiments on the benchmark datasets PASCAL VOC2012 [34] and BSD500 demonstrate that the proposed method provides competitive segmentation results.