A novel method for face recognition

Clustering is such an algorithm which merges the most similar pair of samples into the same classification at every iteration. The traditional similarity evaluation function is manually designed, but the recent interest focuses on supervised or semi-supervised learning where the ground-truth clustered data can be available for training. This paper will first describes how to train a similarity function by regarding it as the action-value function in reinforcement learning. Then, the agglomerative clustering algorithm with superpixel is applied to segment a challenging dataset of brain images. The experimental results demonstrate the proposed method remarkably improved the segmentation accuracy.