Symmetric mean binary pattern-based Pakistan sign language recognition using multiclass support vector machines

Sign languages are a fundamental source of communication for the deaf community, generated through movements of the human body. Similar to the natural languages, sign languages vary from region to region, and from nation to nation. Pakistan sign language (PSL) is inspired by Urdu, national language of Pakistan. It has 38 alphabet signs, and out of them, 36 are represented by the static hand gestures. Automatic recognition of a sign language helps in interaction between the hearing and deaf individuals. The number of quality efforts in context of Pakistan sign language recognition is quite limited, leaving a fair room for addressing open issues of research, i.e., (i) efficient hand detection under complex backgrounds and (ii) extracting signer independent feature vector that should not only be discriminant for all the PSL alphabets but reduced dimension as well. In this research, recognition of PSL static alphabets is addressed where the task of hand localization is accomplished through faster regional-convolutional neural networks (faster R-CNN). Feature extraction is achieved through presenting symmetric mean-based binary patterns (sMBP) that extend uniform local binary patterns. The proposed feature vector not only suppresses the noise but preserves the rotation invariance as well. Classification task is accomplished through error correction output codes (ECOC)-based support vector machines using linear, polynomial and radial basis function kernels with one-vs-one and one-vs-all modalities. The proposed technique is validated through PSL dataset, created by seven native signers, having 7174 images. The comparative results clearly demonstrate the authority of the proposed technique over all of its baseline and competitor techniques.