Using spatial relationships as features in object recognition

Information about the spatial relationships among objects in an image is an important clue towards identifying the objects and in interpreting the contents of the scene. Four primitive spatial relations (LEFT, ABOVE, RIGHT, BELOW) were used as features in a handwritten digit recognition problem. The fuzzy membership values of these four relations between every pair of cavities in each handwritten digit were obtained from the neural network systems and the fuzzy morphology method proposed in the authors' previous work. A feedforward backpropagation network was trained as a digit classifier. Promising results were obtained from the experiments, supporting the concepts that these features are useful for other complex classification problems such as automatic target recognition.