Unsupervised texture-based image segmentation through pattern discovery

This paper presents a new efficient technique for unsupervised segmentation of textured images that aims at incorporating the advantages of supervision for discriminating texture patterns. First, a pattern discovery stage that relies on a clustering algorithm is utilized for determining the texture patterns of a given image based on the outcome of a multichannel Gabor filter bank. Then, a supervised pixel-based classifier trained with the feature vectors associated with those patterns is used to classify every image pixel into one of the sought texture classes, thus yielding the final segmentation. Multi-sized evaluation windows following a top-down approach are utilized during pixel classification in order to improve accuracy both inside and near boundaries of regions of homogeneous texture. Results with synthetic compositions and with complex real images are presented and discussed. The proposed technique is also compared with alternative texture segmentation approaches.

► Supervised classification improves unsupervised segmentation of textured images.
► Classifier effectively trained with samples automatically obtained by clustering.
► Progressive classification with decreasing window sizes improves overall accuracy.