| The nature of automation target recognition is pattern recognition, and feature extraction is an important part of pattern recognition system, which is a powerful guarantee of correct classification. The representative features, which are also invariant to the translation, scaling and rotation, should be select for improving the precision and velocity.Wavelet transform is characterized by multi-resolution, which presents pyramid structure. It coincides well with the way by which people distinguish object from coarse to fineness and from large to tiny. Tow feature extraction methods based on wavelet transform are presented in the dissertation: multiresolution of linear moment, wavelet moment method. The basic principle and realization process of every method are detailed.The approach of linear moment convert the information of 2-D image into 1-D linear moment firstly and then transforms it with multiresolution Orthonormal shell. Feature vector can be get through calculate the power of signal and the coefficients of the signal's FFT in different resolution. This algorithm has been improved for reducing the amount of computing and data and selecting character easily. Wavelet moment is argued. Besides having the invariant to the translation, scaling and rotation, the wavelet moment has the multiresolution properties so it's suitable for classing the very similar objects. Decision and classification is the key problem in target recognition. The minimum-distance and neural network classifier are researched on pattern identification in this paper. Multiresolution analysis combined with series and parallel neural network achieve high recognition rate. Using minimum-distance and neural network classifier together with wavelet moment, we have construct recognition systems witch succeed in classing two class targets on a flat which rotating from degree 0 to degree 360 in the condition of laboratory.
|
PDF Full Text Request