Genetic eigenhand selection for handshape classification based on compact hand extraction

This study proposes compact hand extraction to assist in computerized handshape recognition. First, we devised an image enhancement technique based on singular value decomposition to remove dark backgrounds by reserving the skin color pixels of a hand image. Then, the polynomial approximation YCbCr color model was used to extract the hand. After alignment, we applied lighting compensation to the adaptable singular value decomposition. Finally, a hierarchical pyramid sampling algorithm was used to reduce the impact of variations in handshape. We also constructed a self-eigenhand recognizer with genetic algorithms (GA) for selecting discriminant eigenvector subsets for classification. Although our approach maximizes the differences in hand images for various handshapes, it also minimizes variations in lighting and pose for the same handshape. Experimental results for images from our database and a live sequence showed that our method functioned more efficiently than conventional ones that do not use compact hand extraction against complex scenes. For the 768 images included in inside testing, our classification system achieved an AAR of 99.55% and an FAR of 0.0001%. For live testing, the classification system achieved an accuracy rate of 91.7%, with an error rate of 8.3%. Regarding speed, our system was faster than conventional ones. Our images size was 160 × 120 pixels, operating at an average processing time of less than 1 s per handshape (using an AMD64 Athlon CPU 2.0 GHz personal computer).