Spatial relationships exhibited among regions in an image play an important role in the interpretation of a scene. While humans have an innate ability to recognize spatial relations, it has been difficult to produce algorithms to model these relationships. There have been several attempts to define spatial relationships between objects in a digital image, most recently, with the use of fuzzy set theory. Unfortunately, no method can work effectively in all cases. In this paper, we introduce neural network structures combined with the Choquet fuzzy integration to generalize spatial relationship membership functions. To match the spatial relationship membership value with human intuition, we tested people's perceptions of spatial relations and applied the human test results as the desired outputs of neural networks. Finally, some experimental results on synthetic and real images are shown.