A DTW-based representation to support static sign language recognition from binary images

Abstract

Sign language recognition (SLR) from binary images is a challenging task due to the huge amount of information required to process and the complex variations among classes. Here, we introduced a dynamic time warping (DTW) based representation approach to reveal consistent static SLR patterns from binary images. Indeed, we estimated curvature coefficients sequences (CCSs) from a contour filtering using different step length values. In turn, a DTW feature space is built from CCSs attributes, and a Relief-F-based feature selection stage is carried out to highlight discriminative DTW attributes. Achieved results on a publicly available dataset probe that our strategy attains an 85% classification performance on average. Further, to the best of our knowledge, this is the first attempt to apply dissimilarity-based representations for codifying binary images in SSLR.