Based on an analysis of the wavelet multi-scale transform in target image, this paper puts forward a new method to extract the main features of the transform. It, in terms of the energy concentricity of the image's wavelet coefficients, selected the parts with concentrated energy to construct the feature vectors, which include most of the edge, texture, luminance and structure features. As there is no limit to the distribution of the image feature information, the method can be used in many kinds of image feature extraction, thus solving the problem of feature illegibility with which the single-feature extraction is confronted. In the experiment, normal white noises with different ranges were added to the images and the result approves that the feature vectors are robust to noise.