Feature learning-based segmentation algorithm for hand segmentation

Feature learning algorithms have been studied extensively for solving many pattern recognition problems, and several effective algorithms have been proposed. This paper proposes a feature learning-based segmentation algorithm (FLSA) for determining appropriate features for hand segmentation. This new approach combines an unsupervised learning phase and a supervised learning phase for hand segmentation. The unsupervised learning phase consists of preprocessing, patch generation, and filter learning through the K-means algorithm. The supervised learning phase consists of feature extraction, classification learning, pixel classification, and morphological operation. The FLSA starts feature learning through the K-means algorithm and then trains a neural network (NN) as the classifier for classifying a pixel into two categories: hand and nonhand. As feature learning progresses, features appropriate for hand segmentation are gradually learned. The emphasis of the FLSA on feature learning can improve the performance of hand segmentation. The Georgia Tech Egocentric Activities data set was used as unlabeled data for feature learning, and the corresponding ground-truth data were used for NN training. The experimental results show that the FLSA can learn features from unlabeled data and verify that feature learning leads to hand segmentation that is more effective than that without feature learning.