Image Classification by Lifting Wavelet Pca

An image classification method based on the lifting wavelet and PCA is proposed. First, several training images chosen from given classes are decomposed into low-pass and high-pass images by using wavelet transform. Applying PCA to all the low-pass images, principal component vectors are computed. The feature vectors of low-pass images are constructed by expanding the low-pass images with respect to the principal component vectors. The average of the obtained feature vectors in each class is calculated, and lifting parameters are learned so that the lifting feature vectors in each class approach to the average in the same class. Lifting low-pass images for training images are computed exploiting the learned parameters. PCA is applied again to these images for improving the feature vectors of training images. This process is repeated until the classes are separated sufficiently. Classification of a query image is accomplished by comparing its lifting feature vector with the feature vectors for training images. The validity of our method is checked using a benchmark data and object images captured by a robot camera.