Unsupervised image segmentation via Stacked Denoising Auto-encoder and hierarchical patch indexing

For decades, image segmentation is a hot research direction in computer vision because of its extensive and practical applications. In this work, we propose a method of image segmentation based on auto-encoders and hierarchical clustering algorithm, aiming at dealing with the segmentation problem in an unsupervised way. More specifically, this proposed method consists of two stages: training and segmenting. In training stage, we divide sample images into non-overlapped patches and extract deep-level feature representations from the patches using Stacked Denoising Auto-encoder (SDA), then we perform unsupervised and hierarchical K-means clustering on these feature representations and build an indexing tree structure. In segmenting stage, we achieve segmentation of an arbitrary image based on the indexing tree structure. This unsupervised methodology is demonstrated to be an improvement over traditional unsupervised segmentation methods owing to the introduction of sample images. Experimental evaluations on several benchmark datasets indicate that our algorithm outperforms several other methods in both time efficiency and accuracy.