A fuzzy neural network system suitable for image analysis is proposed. Each neuron is connected to a windowed area of neurons in the previous layer. The operations involved follow a method for representing and manipulating fuzzy sets, called Composite Calculus. The local features extracted by the consecutive layers are combined in the output layer in order to separate the output neurons in groups in a self-organizing manner by minimizing the fuzziness of the output layer. In this paper we focalize our attention on the application of the proposed model to the edge detection based segmentation, reporting results on real images and investigating the robustness of the system with noisy data.