Composited Sketch Recognition Via String Kernel

Sketch recognition is one of the essential step of sketch understanding. As anstructural data, the inner structure of sketch can convey plentiful information. However,in current approaches to sketch recognition, sketch representations are transformed intoa feature vector description, lose all the structure information.Unlike vectors, strings are structural data, providing information about the struc-ture of an object. String kernel is defined on string, can handle the structural informa-tion of objects. String kernel has been applied successfully in the domain of naturelanguage process and biological information detector. Composition of a natural sketchscene is an analog to the composition of a text, that is, sentences are built up fromphrases, phrases are built up from alphabets. Sketch can also be represented by a groupof predefined primitives. So, this thesis examined if a structured description of sketch,namely feature strings, can be used for sketch recognition. This thesis introduces theapplications of string kernel in sketch understanding. A sketch recognition prototypesystem via string kernel is implemented.The main contents can be summarized as follows:1. Learning algorithm based on string kernel: String kernel has been successfullyapplied in the field of text classification, According to the fact that sketch under-standing have an analogy with text classification, we introduce string kernel intothe file of sketch understanding.2. Spatial relationship is extracted to represent the complex sketch symbols, to traina hierarchical two-stage string kernel SVM classifier. Experiments on digitalsketch shows that sketch recognition via string kernel can achieve good perfor-mance.