Embedded Files
Projeto: Reconhecimento de Gestos Estáticos aplicados a Linguagem Brasileira de Sinais (Libras)
Igor L. O. Bastos
Michele F. Angelo
Angelo C. Loula
This work describes the development of an approach for recognition of static gestures in images applied to Brazilian Sign Language (Libras) as a practical context. This language is used by deaf people in Brazil and is considered, since 2002, as an official Brazilian language. Our gesture recognition approach lies on the combination of two shape descriptors, image processing techniques and a two-stage neural network classifier to recognize the gestures.
The proposed approach advances, when compared to related work, in terms of amount of gestures addressed, (higher than most of gesture recognition studies, especially those related to Libras); and for intending to make the dataset of images publicly available, which could aid the formulation of other studies and comparison of the present approach with other proposals.
In addition, we also developed an strategy for skin segmentation that can be understood as a contribution of this work, being able to be applied in contexts beyond the gesture recognition. In this work the signs are recognized statically, representing only one Libras parameter (hand configuration). However, Libras signs also include body language, movement and orientation.
Image dataset
Images from different Libras signs were used to compose the dataset. Three Libras experts and two deaf students volunteered to be models to compose the dataset, which contains 9600 images.
The set of signs is composed by: (i) letters from the Libras alphabet: A, B, C, D, E, F, G, I, L, M, N, O, P, Q, R, S, T, U, V, W, X, Y, (ii) numbers: 1, 2, 4, 5, 7 and 9; and (iii) words: Plane, Word, Adult, America, House, Gas, Law, Identity, Together, Stone, Little and Verb, totaling 40 different signs.
To select the signs that would be recognized in this work, Libras experts were asked to select those signs recognizable by only the hand configuration parameter. Signs that have other parameters, like distance to certain parts of body or hand movements, were not considered except those in which the hand configuration is sufficient to allow the recognition.
For each one of the aforementioned signs, 240 images (with resolution of 50x50 pixels) were acquired, totaling 9600 images. Half of them are grayscale images representing the gestures. The other half corresponds to binary masks obtained with a skin detection approach (applied before converting images to grayscale) and represents the skin zones of the 4800 images obtained before, as shown on Fig. 1.
The acquisition process was performed considering a standard distance from the camera to the individuals used as models. In addition, the images were acquired considering some small variations on lightning and a simple (white) background. The image dataset also considers some variations in terms of hand postures and hand sizes, which are particular to each individual, as shown on Fig. 2.
The dataset is available for download. The dataset is free to use for scientific research. The citation paper for this dataset is
I. L. O. Bastos, M. F. Angelo and A. C. Loula, "Recognition of Static Gestures Applied to Brazilian Sign Language (Libras)," 2015 28th SIBGRAPI Conference on Graphics, Patterns and Images, Salvador, 2015, pp. 305-312.
doi: 10.1109/SIBGRAPI.2015.26
Text available at IEEE and SIBGRAPI Digital Library Archive.
If you use this dataset, cite the paper reference, not this website.
Images and binary masks are separated in folders corresponding to 6 folds. The zip file size is 19,5 MB. DOWNLOAD HERE.
Page updated
Report abuse