BdSLW-11: Dataset of Bangladeshi sign language words for recognizing 11 daily useful BdSL words

The dataset of Bangladeshi sign language words (BdSLW) is rare. Though there are lots of datasets of BdSL sign alphabets, numbers, or characters, there are not enough datasets of sign words. This is the first dataset about sign words of BdSL according to the author(s) knowledge. So, this dataset is developed by collecting data from people. This is an image dataset. This dataset is a collection of 1105 images of sign words. A total of 11 sign word categories are selected which are important and daily use in our life. As this is an image dataset, so the images of sign words are taken by camera from the sign users of Bangladesh. Authors have gone to the individuals of sign users and captured images from them with their permission. Then the images are analyzed and segmented into the images which have quality such as no background, clear, bright, etc. This dataset is used for recognizing BdSL sign words.


a b s t r a c t
The dataset of Bangladeshi sign language words (BdSLW) is rare.Though there are lots of datasets of BdSL sign alphabets, numbers, or characters, there are not enough datasets of sign words.This is the first dataset about sign words of BdSL according to the author(s) knowledge.So, this dataset is developed by collecting data from people.This is an image dataset.This dataset is a collection of 1105 images of sign words.A total of 11 sign word categories are selected which are important and daily use in our life.As this is an image dataset, so the images of sign words are taken by camera from the sign users of Bangladesh.Authors have gone to the individuals of sign users and captured images from them with their permission.Then the images are analyzed and segmented into the images which have quality such as no background, clear, bright, etc.This dataset is used for recognizing BdSL sign words.

Value of the Data
• This dataset is greatly useful for the implementation of both image processing and image classification.• These data are useful because they will be used to research sign language words recognition automatically.
The Deaf and Dumb (D&D) community will greatly benefit as these data will be used by researchers to develop automatic sign language words recognition.• The researcher of Computer Vision can use these data for further experiments to achieve better performance in recognizing sign language words.• The dataset is essential for recognizing Bangladeshi Sign Language (BdSL) Words.

Objective
The main objective behind the generation of the BdSLW-11 dataset is to contribute to researchers developing a system to remove the communication barrier of the Deaf and Dumb (D&D) community of Bangladesh.This dataset will help researchers greatly to do their research in Bangladeshi Sign Language (BdSL) especially sign words.We have developed this dataset.This dataset supports an original research article that has been published.So, this data article adds value to the published research article by demonstrating the dataset functionality clearly and its effectiveness in recognizing BdSL sign words.This article will comprehensibly give an idea to readers about the BdSLW-11 dataset.This will help them to easily understand the dataset and reuse the dataset if they want to do further research in BdSL sign words.

Data Description
Sign language is different from normal languages.The Deaf and Dumb (D&D) community use this language.They use various gestures to communicate among themselves and also with normal people.Sign language is dynamic or static hand gestures [2] .D&D community suffers unexplained sufferings for communication.Computer vision can greatly contribute to eradicating their suffering.Com puter vision is being used for various real-life problems [3] .Researchers are developing various models or systems using computer vision to recognize hand signs and translate them into voice or text [4] .These systems or models require a dataset to experiment with.Though there are a lot of Bangladeshi Sign Alphabets, Numbers, and characters datasets available [5][6][7] , there is no Bangladeshi Sign Words Dataset available publicly.
In this work, we have developed a Bangladeshi sign language words (BdSLW) dataset for recognizing 11 daily useful common sign words.The selected sign words are 'Bad', 'Beautiful', 'Friend', 'Good', 'House', 'Me', 'My', 'Request', 'Skin', 'Urine' and 'You'.The images are processed as cropped and separated as qualified based on the focus on hand gestures, clear background, and brightness.This dataset supports an original research article [1] .Fig. 1 describes the methodology of constructing the BdSLW-11 dataset and Fig. 2 shows the total number of images in each class.
The hand sign of the 11 sign words is taken manually from real volunteer signers with their full permission.We have captured images of hung signs when the signer performs the sign.The images are then analyzed and removed from the other human body except for the hand portion.Images were resized as 224 × 224 pixels and filtered the images.If the images were qualified, they are stored in the dataset as their class.Otherwise, the images were reanalyzed until fit for use.curacy.So, there are not enough variations and lighting.Though BdSLW-11 has some limitations, this is the first Bangladeshi sign language words dataset as per author(s) knowledge.

Experimental Design, Materials and Methods
The proposed methods of recognizing the Bangladeshi sign language words (BdSLW) consist of two phases.One is data augmentation and the other is 11 classes of BdSLW recognition.The method is based on capturing the image of performed sign of the user and using the trained transfer learning (TL) models to classify.Finally, it displays the recognized BdSL sign words of the input sign as Fig. 3 .
The dataset is evaluated by deep learning models especially transfer learning (TL) techniques to recognize the 11 BdSL sign words.The most popular transfer learning techniques are VGG16, VGG19, InceptionV3, AlexNet, ResNet-50, Xception, and DenseNet [4] .We have used VGG16, VGG19, InceptionV3, ResNet-50, and AlexNet to train our dataset for evaluation.The accuracy of the employed TL methods that have been trained on the BdSLW-11 dataset is depicted in Fig. 4 .
Fig. 4 demonstrates the accuracy of the five pre-trained models of transfer learning techniques that have been trained on our developed BdSLW-11 sign words recognition dataset.The models have achieved 99.92%, 99.58%, 98.70%, 68.66% and 97.86% accuracy for VGG16, VGG19, InceptionV3, ResNet50 and AlexNet respectively.This shows that the VGG16 model achieves the Dataset size greatly impacts the accuracy and recognition of sign language.So before training the transfer learning techniques on our dataset, it is recommended to augment the dataset.This will enhance the size of the dataset.The implemented TL models of the above accuracy have been done on the augmented dataset.The augmented dataset size was 3835 images of the training set.We had split the 1105 images into 850 images for the training set and 255 images for the testing set.Augmentation is various techniques to enlarge the dataset such as flipping, color space, cropping, rotation, translation, noise injection, etc. [8] .We have used flipping, rotation, and zooming augmentation techniques to augment the dataset which enhances the size of the dataset.The augmentation code is as follows:

Fig. 2 .
Fig. 2. Image count of each class in the dataset.

Fig. 4 .
Fig. 4. Comparison of different TL model accuracy on BdSLW-11 dataset.best accuracy.VGG16 TL model performs better to recognize 11 daily useful sign words comparing other models trained on our BdSLW-11 dataset.Dataset size greatly impacts the accuracy and recognition of sign language.So before training the transfer learning techniques on our dataset, it is recommended to augment the dataset.This will enhance the size of the dataset.The implemented TL models of the above accuracy have been done on the augmented dataset.The augmented dataset size was 3835 images of the training set.We had split the 1105 images into 850 images for the training set and 255 images for the testing set.Augmentation is various techniques to enlarge the dataset such as flipping, color space, cropping, rotation, translation, noise injection, etc.[8] .We have used flipping, rotation, and zooming augmentation techniques to augment the dataset which enhances the size of the dataset.The augmentation code is as follows: