Abstract:

The core of the k nearest-neighbor classification method was put forward to search for the k nearest-neighbors of a new sample (feature vector). The important information hiding in the Haar wavelet coefficients was investigated. Then the relationship between the Haar wavelet coefficient and the variance of a vector was determined, from which an inequality about the wavelet coefficient was obtained. When searching for the k nearest-neighbors  this inequality condition was employed to identify and kick out quickly many vectors that are impossible to be the k closest vectors in the design set; thus the k nearestneighbors can be found quickly. The k nearest-neighbor classification method with this fast algorithm can save substantially the classification time; meanwhile achieve the same classification performance as that with the exhaust search algorithm. Experiments on texture classification are performed and the results validate the proposed algorithm.

Key words: communication, signal processing, wavelet transform, k nearestneighbor classifier, texture classificationk