Unsupervised Learning of Image Segmentation Based on Differentiable Feature Clustering

Abstract

The usage of convolutional neural networks (CNNs) for unsupervised image\nsegmentation was investigated in this study. In the proposed approach, label\nprediction and network parameter learning are alternately iterated to meet the\nfollowing criteria: (a) pixels of similar features should be assigned the same\nlabel, (b) spatially continuous pixels should be assigned the same label, and\n(c) the number of unique labels should be large. Although these criteria are\nincompatible, the proposed approach minimizes the combination of similarity\nloss and spatial continuity loss to find a plausible solution of label\nassignment that balances the aforementioned criteria well. The contributions of\nthis study are four-fold. First, we propose a novel end-to-end network of\nunsupervised image segmentation that consists of normalization and an argmax\nfunction for differentiable clustering. Second, we introduce a spatial\ncontinuity loss function that mitigates the limitations of fixed segment\nboundaries possessed by previous work. Third, we present an extension of the\nproposed method for segmentation with scribbles as user input, which showed\nbetter accuracy than existing methods while maintaining efficiency. Finally, we\nintroduce another extension of the proposed method: unseen image segmentation\nby using networks pre-trained with a few reference images without re-training\nthe networks. The effectiveness of the proposed approach was examined on\nseveral benchmark datasets of image segmentation.\n