Improvement of Supervised Shape Retrieval byLearning the Manifold Space

Manifold learning is the technique that aims forfinding a constructive way to embed the data from a highdimensionalspace into a low- dimensional manifold based on nonlinearapproaches. In this paper a supervised manifold learningmethod for shape recognition is proposed. The approach isbased on learning the manifold space for training samples andmap the test samples to the learned space by a GeneralisedRegression Neural Network (GRNN). The main goal in this paperis to propose a new feature vector to coincide semantic andEuclidean distances. To accomplish this, the desired topologicalmanifold was learnt by a global distance driven non-linear featureextraction method. The experiments showed that the geometricaldistances between the test samples on the manifold space aremore related to their semantic distance. To fuse the results ofshape recognition based on contour and region based methods, inour framework the final result of shape recognition is based oncommittee decision in three manifold spaces. The experimentalresults confirmed the effectiveness and validity of the proposedmethod.