Intelligent image analysis using adaptive resource-allocating network

This paper presents a unified image analysis approach for object detection, segmentation, and classification using an adaptive resource-allocating network (ARAN), which is based on using unsupervised learning to cluster shapes and supervised learning to classify objects. The proposed neural network is incrementally grown by adjusting the clusters, and by creating a new cluster whenever an unusual shape is presented. Each hidden node represents a cluster, with centers and widths of the hidden nodes used as templates to provide faster and more accurate object detection and segmentation. On-line learning gives the system improved performance with continued use. The effectiveness of the resulting system is demonstrated on the task of diagnosing breast cancer.