Recognition of a real-time signer-independent static Farsi sign language based on fourier coefficients amplitude

A sign language is a language which uses manual communication and body language to convey meaning, as opposed to acoustically conveyed sound patterns. This can involve simultaneously combining hand shapes, orientation and movement of the hands, arms or body, and facial expressions to fluidly express a speaker's thoughts. Sign language is a preliminary communication way for individuals with hearing and speech problems. Considering that more than a hundred million people all around the world are annoyed by hearing loss and impaired speech, it is needed to design a system for automatic sign language interpreter as an interface between deaf-and-dumb and ordinary people can feel it strongly. Given this, in this article we aimed to design such a computer vision-based translation interface. Farsi sign language recognition is one of the most challenging fields of study is given because of some features such as the wide range of similar gestures, hands orientation, complicated background, and ambient light variation. A Farsi sign language recognition system is based on computer vision which is capable of real-time gesture processing and is independent of the signer. Furthermore, there is no need to use glove or marker by the signer in the proposed system. After hand segmentation from video frames, the proposed algorithm extracts boundary of the dominant hand to compute the normalized accumulation angle and represents the boundary, so that the invariance to transition and scale change of the features is realized at this stage. Afterward, Fourier coefficients amplitude is extracted as preferred features at the frequency domain, while invariance to rotation of the features is added at this point. Then the frequency features, as extracted features for gesture recognition, are applied to inputs of feed-forward multilayer perception neural network. The proposed method is presented to make recognition system independent of the signer and retrofit it against signer's distance changes from camera using features of powerful invariant extraction against transition, scale change, and rotation. Data classification is carried out by three classifiers including Bayes, K-NN, and neural network. Performance of the classifiers was also compared. Training set of gestures comprised 250 samples for 10 gestures and 5 positions and orientations that were performed by 5 individuals. Recognition results showed an outstanding recognition rate of the system.