Features that capture well the textural patterns of a certain class of imagesare crucial for the performance of texture segmentation methods. The manualselection of features or designing new ones can be a tedious task. Therefore,it is desirable to automatically adapt the features to a certain image or classof images. Typically, this requires a large set of training images with similartextures and ground truth segmentation. In this work, we propose a framework tolearn features for texture segmentation when no such training data isavailable. The cost function for our learning process is constructed to match acommonly used segmentation model, the piecewise constant Mumford-Shah model.This means that the features are learned such that they provide anapproximately piecewise constant feature image with a small jump set. Based onthis idea, we develop a two-stage algorithm which first learns suitableconvolutional features and then performs a segmentation. We note that thefeatures can be learned from a small set of images, from a single image, oreven from image patches. The proposed method achieves a competitive rank in thePrague texture segmentation benchmark, and it is effective for segmentinghistological images.
展开▼
