Shape recognition using complex nonlinear exponential autoregressive model

In this paper, the closed boundary of an arbitrary 2-D shape is considered to be physically related to the trace of a 2-D orthogonal nonlinear vibration with equal period, and hence a complex exponential autoregressive (CEAR) model is proposed to describe the 2-D closed boundary. The model coefficients are invariant to translation, rotation, scale and choice of the starting point in tracing a boundary, additionally, they are not invariant to mirroring transformation. Due to the nonlinearity of the model, the local information of boundary is also reflected in the coefficients. Experimental results indicate that this model has superior performance in recognizing similar shapes and some different patterns with mirroring similarity. Furthermore, this model has good prospect for the recognition of constrained handwritten numerals and characters