Object recognition by sub-scene graph matching
Wen-Jing Li, Tong Lee · 2002
In this paper a new method, called sub-scene graph matching, for object recognition by Hopfield neural network is proposed. The method divides a scene graph into many sub-scene graphs, and a second order Hopfield neural network is constructed to obtain local matches between each sub-scene graph and the model graph. Finally, we can achieve a correct match for the whole scene graph from the statistics of the local matches. Experimental results demonstrate that the presented method can efficiently recognize isolated or occluded 2D objects invariant to translation, rotation, and scale, even when the objects are distorted in shape, provided the local structures are unchanged.