A Graph-Based Active Learning Approach Using Forest Classifier for Image Retrieval

Content-Based Image Retrieval System is comprised of large image collections which find their use in applications such as statistical analysis, medical diagnosis, photograph archiving, crime prevention, face detection, etc. This poses a challenge for pattern recognition techniques used in image retrieval, which require being both efficient and effective. These techniques involve high computational burden in training phase, to separate samples from distinct classes, for active learning. In active learning paradigm, system first returns a small image set. Inference is drawn from this image set based on relevance as per user perception. Hence image retrieval based on context suffers in terms of precision and recall, as retraining and interactive time response is involved. A classifier known as Optimum-Path Forest (OPF) reduces this computational overhead by transforming the problem of classification as a graph obtained from dataset samples in feature space. It involves fast computation of trees built through the forest classifier in a graph resulting from training dataset samples. An optimum path is conquered from least cost approximation of a shortest path through the current prototype sample to the image under consideration.