Efficient image classification using neural networks and multiresolution analysis

The authors investigate a new efficient image classification strategy. They propose a multiresolution analysis of the images to be classified and use of feedforward neural networks to classify the images at various lower resolutions. This approach results in a major reduction of the networks´ interconnection weights as well as the required learning times. The proposed approach is applied first to the images of the lowest resolution; if the classification results are not acceptable, it is successively repeated to the next images of higher resolution. A neural network architecture which incorporates most of the interconnection weights already computed at the lower level (i.e., the knowledge already acquired by the network of the previous resolution level) is proposed for this purpose. Experimental results illustrate the efficiency of the proposed multiresolution classification procedure in a real life application.<>