Abstract: To speed up the modernization of stockbreeding and overcome the defects of the low accuracy of individual dairy cattle recognition with traditional methods, the traditional KPCA (kernel principal component analysis) method was improved from two angles of reducing the covariance matrix dimension and introducing category information. The research of combining wavelet transform with improved KPCA was applied for recognition dairy cattle based on the texture feature. Firstly, the normalized dairy cattle image was decomposed by wavelet transform to obtain four sub-graphs. Then an improved KPCA algorithm was used for feature extraction of each sub-graph and the feature matrix was obtained by weighting the feature components. Finally, multi-class SVM algorithm was built for training and classification. The pre-collected 20 dairy cattle's videos were converted into image sequence and 20 000 images were chosen to form experiment data sets. Through several groups of experiments for three important parameters of the wavelet fusion weights, the number of fusion vector groups and the feature dimension, values were set and then experiments of individual dairy cattle recognition were performed using different algorithms. The results showed that it took only 4.20 s for the proposed method reach the accuracy of 96.31%, which has obvious advantages over other algorithms. It can be appropriately applied to the field of dairy cattle individual recognition with high performance and low cost.

Key words: wavelet transform, improved KPCA, feature fusion, dairy cattle, individual recognition