Multi Feature Small Sample Object Recognition Method based on DTW Algorithm

Abstract

In the process of small sample object identification, the use of single feature or multi-featured one-dimensional data classification methods has the disadvantages of low accuracy and large amount of sample data. A multi-feature-based small sample recognition method based on DTW algorithm is proposed. Selecting four kinds of object features and using dynamic time warping (DTW) algorithm to regularize the feature of the selected object boundary chain code, to obtain the same dimension of the chain code data for the similarity calculation. Finally, using the Probabilistic Neural Network (PNN) to perform the object recognition training on the four normalized features separately and in different combinations, the features with strong correlations for object type identification are screened out. Compared with K-nearest neighbor algorithm (KNN) recognition results, the method can reduce the data dimension from 11025 to 18, and the recognition accuracy rate is up to 93.9%.