T.T. Nguyen, V.G. Spitsyn
In the recent years hand recognition has attracted interest of many researchers. Different algorithms and methods have been proposed to archive high accurate and robust recognition of hand and gestures. This paper proposes a new algorithm for solving the hand recognition problem using the SURF features and neural network. Since introduction, SURF (Speeded-up Robust Feature) has gained its popularity in image categorization, image indexing and retrieval for its robustness and performance. However, in object recognition (and hand recognition particularly) SURF features cannot be used directly in neural network (or other classifiers) for its variability in length. In this paper a new method for creating descriptors for neural network classifier based on SURF is introduced. The extracted SURF features are clustered using K-means clustering to create a “vocabulary” of visual words. The SURF feature set extracted from an image is then compared with the vocabulary as a look-up table to choose appropriate visual words that correctly describe the image SURF feature set. A histogram of occurred words in that image is built and will be used as the input image descriptor for recognition in a multilayer neural network with back-propagation. The experiments have proved that the proposed algorithm can work well in real time at the processing speed of 15 frames per second with average accuracy of 92%. Based on the proposed algorithm a program packet for real-time video hand recognition has been built and introduced.