Combination of Prior Relations with Visual Information by Random Walk Model for Object Recognition

Traditional object recognition methods in computer vision are almost based on just the visual features, which cannot perform well in a more complex circumstance. To attack this critical problem, this paper proposes a novel object recognition method which combines object recognition with the prior relations. During the training stage, structured presentation of the prior relations is applied through a hybrid graph which contains image similar sub-graph, semantic similar sub-graph and the relations between the two sub-graphs. A random walk model is then constructed according to the hybrid graph. During the recognition stage, a new testing image node is added to the random walk model. Then the relations between this node and the nodes in the random walk model are calculated. At last, random walks which start from the testing image node are performed at the random walk model. The probability rank provided by the result of random walks will serve as the recognition result of the testing image. Experimental results illustrate the validity and stronger recognition performance of the proposed method.