This article presents a feed-forward network architecture that can be used as a nonlinear feature extractor for texture segmentation. It comprises two layers of feature extraction units; each layer is arranged into several planes, called feature maps. The features extracted from the second layer are used as the final texture features. The feature maps are characterised by a set of masks (or weights), which are shared among all the units of a single feature map. Combining the nonlinear feature extractor with a classifier, we have developed a texture segmentation system that does not rely on pre-defined filters for feature extraction; the weights of the feature maps are found during a supervised learning stage. Tested on the Brodatz texture images, the proposed texture segmentation system achieves better classification accuracy than some of the most popular texture segmentation approaches.