Generation of Hand Contour Images using Colored Glove and Its Application to Hand Pattern Recognition

Abstract

In this paper, we show the effectiveness to use not only canny edge images but also hand contour images for hand pattern recognition by convolutional neural network (CNN). Hand contour images are binary images with only hand shapes and fingers information including fingers curled information. These hand contour images are generated using the colored glove which we proposed in the previous studies. In the experiments of hand pattern recognition, we investigate recognition accuracy by cross-validation method. Learning model using CNN consists of 4 convolution layers and 4 pooling layers. Moreover, Network In Network (NIN) is adopted as a convolution method. Test dataset is composed of only canny edge images, taking account of applicability for sign language recognition and so on. On the other hand, training dataset is composed of combination of canny edge images and hand contour images. Through the recognition experiments, we confirm the effectiveness of combining hand contour images in training dataset. The highest average recognition accuracy is 96.2% when combination rate of canny edge images and hand contour images is 50:50. This value is 6.9% higher, compared with the case combination rate is 100:0.