Amharic Sign Language Data Sets

Hearing-impaired people use Sign Language to communicate with each other as well as with other communities. Usually, they are unable to communicate with normal people. Most of the people without hearing disability do not understand the Sign Language and unable to understand hearing-impaired people. So, they need recognition of Sign Language to text. A data set is collected from different teachers of sign language for recognition of Amharic Sign Language to Amharic characters. After gathering data with different backgrounds and positions of hands, the data is prepared for feeding to machine learning. The initial step for preprocessing data is frame extraction of video followed by resizing data and feature extraction. By using the LabelImg tool we annotated this frame of data to create XML, TFrecord, and CSV. Finally, we fed the data into two different models: Faster R-CNN and Single-Shot Multibox (SSD) detector. From the result of the models, we justified for Amharic Sign Language recognition to characters the SSD better in accuracy than Faster R-CNN but Faster RCNN is good with accuracy. Anyone can use the data for recognizing Amharic Sign Language to Amharic characters from the image, video, and Real-time. We used 10 class for recognizing the characters, the research will continue to include the remaining six orders, words, and sentences used in Sign Language to have a full-fledged Sign Language recognition model to a complete system.听力受损人群借助手语实现彼此间及跨社群的交流，但通常难以与健听人群沟通。多数非听障人士不懂手语，无法理解听障群体的表达，因此亟需手语转文本的识别技术。本数据集采集自多位阿姆哈拉语手语（Amharic Sign Language）教师，用于实现阿姆哈拉语手语到阿姆哈拉文字符的识别任务。在收集涵盖不同手部姿态与位置的样本后，对数据进行预处理以适配机器学习模型输入：首先从视频中提取帧，随后调整数据尺寸并提取特征。我们使用LabelImg工具对这些帧数据进行标注，生成XML、TFrecord及CSV格式的标注文件。最终将预处理好的数据输入两种不同模型：Faster R-CNN与单发多框检测器（Single-Shot Multibox Detector, SSD）。从模型测试结果来看，在阿姆哈拉语手语转文字符识别任务中，SSD的准确率优于Faster R-CNN，不过Faster R-CNN同样具备不错的准确率表现。任何使用者均可利用该数据集，实现从图像、视频及实时场景中的阿姆哈拉语手语到阿姆哈拉文字符的识别。本研究共使用10个字符类别完成识别任务，后续将拓展涵盖手语中剩余的6类单元、词汇与语句，以构建完整的手语识别全流程系统。