Recombining Features of Frequency Domain and Location for Machine Recognition of Sign Language

In this article, a system for recognizing Persian sign language alphabets is presented. This system is able to recognize 32 hand postures for Persian alphabets and translate it into Persian text. For this purpose, images of hand positions have been considered for each letter of the alphabet. The database contains 600 images of different people taken by a digital camera. We have transferred all the image data to the binary domain and resized them with a single scale. Image data preprocessing includes image cropping and noise removal. After pre-processing, 3 algorithms are proposed to extract features. The proposed algorithms include the image segmentation algorithm, the distance between border contour points and the center of gravity algorithm, and Radon transformation. Algorithm of the distances between the border contour points and the center of gravity shows how the points are placed on the peripheral curve of the hand in relation to each other and to the center of gravity, and therefore provides suitable structural information for describing states. The next algorithm is based on image segmentation. In this algorithm, the ratio of the number of white pixels to the total number of pixels is calculated in each of the areas. In Radon transformation, in addition to obtaining the general information of the image in each of the modes, we have increased the accuracy of the detection by using the proposed method and discarding additional information. The proposed methods also provided good results on other image databases.