Composite Sketch Shape Recognition Based On Support Vector Machines

Sketch shape recognition is the basis of sketch semantic understanding, and it consists of two sequential and cyclic phases: primitive sketch shape recognition and composite sketch shape recognition. The former got a lot of research and progress, while the later got less attention.Support Vector Machines (SVM) is a system for efficiently training the linear classification machines in the kernel-induced feature spaces, with underlying generalization theory and optimization theory. However, there are few work of composite sketch recognition based on SVM, and the existing works have not taken full advantage of SVM's powerful capability of learning, classification and generalization.The thesis brings forward a composite shape description method based on the space constraints of primitive shapes, and a DAGSVM based recognition algorithm is proposed correspondingly. The approach makes use of time sequence information of strokes, and integrates the advantages of feature-based and stroke-based recognition approach. The main contents can be summarized as follows:1) The state-of-arts of sketch recognition are reviewed first in the thesis. We put forward a new composite sketch shape description method based on space constraints of primitive shapes.2) We design a heuristic algorithm of parameter search with the law of RBF kernel function. The algorithm enhances the efficiency of grid search and the parameter scope is reduced.3) A composite shape recognition algorithm based on Directed Acyclic Graphs SVM (DAGSVM) is proposed.4) The experiments have showed the feasibility of the new method and the recognition algorithm.