Retrainable neural networks for image analysis and classification

A novel approach is presented for improving the performance of neural network classifiers in image recognition, segmentation or coding applications, based on a retraining procedure at the user level. The procedure includes: a maximum a posteriori (MAP) estimation technique for optimally selecting a retraining data set from the image applied to the network during real life operation; a decision mechanism for automatic activation of network retraining; and a neural network module which performs the classification task. The extracted feature set, used for retraining the network, can include additional elements compared to those used in the network initial training phase, so that it better fits the specific application data under consideration. Results are presented which illustrate the theoretical developments as well as the performance of the proposed approach in real life experiments.