The paper presents a road junction operator, which was developed for medium resolution black-and-white orthoimages. The operator uses a feed-forward neural network applied for a running window to decide whether it contains a 3- or 4-arm road junction or not. The training set was created by a data analysis based feature selection. The best features took part in the training of 3-layer neural networks. The features are coming from the central kernel of the window (raster data) and from edge detection (vector data). The vectorized edges are only kept for training, if they are going through the central circle, which represents the junction central in a rotation invariant way. The edges fulfilling the circle criterion are applied to derive features, like edge length and direction measures. A set of identically structured networks with varied parameters was generated and trained by an efficient optimization algorithm. The evaluation of the networks was carried out in in-sample tests, where the main traditional methods are compared to the neural solution. The out-of-sample test was performed by real image chips with different rotations. The obtained results demonstrate the principal feasibility of the developed method.