Scientific Publications Directory

Full Author List

Hayati; Lilis Nur, Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Malang, Indonesia; Department of Computer Science, Universitas Muslim Indonesia, Makassar, Indonesia; Handayani, Anik Nur; Irianto, Wahyu Sakti Gunawan, Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Malang, Indonesia; Asmara, Rosa Andrie, Information Technology Department, Politeknik Negeri Malang, Malang, Indonesia; Indra, Dolly, Department of Computer Science, Universitas Muslim Indonesia, Makassar, Indonesia; Damanhuri, Nor Salwa, Electrical Engineering Studies, Universiti Teknologi MARA (UiTM),Cawangan Pulau Pinang, Malaysia

Abstract

Sign language recognition plays a critical role in promoting inclusive education, particularly for deaf children in Indonesia. However, many existing systems struggle with real-time performance and sensitivity to lighting variations, limiting their applicability in real-world settings. This study addresses these issues by optimizing a BISINDO (Bahasa Isyarat Indonesia) alphabet recognition system using the SSD MobileNetV2 architecture, enhanced with gamma correction as a luminance normalization technique. The research contribution is the integration of gamma correction preprocessing with SSD MobileNetV2, tailored for BISINDO and implemented on a low-cost assistive robot platform. This approach aims to improve robustness under diverse lighting conditions while maintaining real-time capability without the use of specialized sensors or wearables. The proposed method involves data collection, image augmentation, gamma correction (γ = 1.2, 1.5, and 2.0), and training using the SSD MobileNetV2 FPNLite 320x320 model. The dataset consists of 1,820 original images expanded to 5,096 via augmentation, with 26 BISINDO alphabet classes. The system was evaluated under indoor and outdoor conditions. Experimental results showed significant improvements with gamma correction. Indoor accuracy increased from 94.47% to 97.33%, precision from 91.30% to 95.23%, and recall from 97.87% to 99.57%. Outdoor accuracy improved from 93.80% to 97.30%, with precision rising from 90.33% to 94.73%, and recall reaching 100%. In conclusion, the proposed system offers a reliable, real-time solution for BISINDO recognition in low-resource educational environments. Future work includes the recognition of two-handed gestures and integration with natural language processing for enhanced contextual understanding.

Related SDGs Targets

Full Author List

Hayati, Lilis Nur; Indra, Dolly; Fahmi, Universitas Muslim Indonesia; Handayani, Anik Nur; Irianto, Wahyu Sakti Gunawan, Universitas Negeri Malang; Asmara, Rosa Andrie, Politeknik Negeri Malang

Abstract

Indonesian Sign Language (BISINDO) is one of the sign languages used in Indonesia. The process of classifying BISINDO can be done by utilizing advances in computer technology such as deep learning. The use of the BISINDO letter classification system with the application of the MobileNet V2 FPNLite SSD model using the TensorFlow object detection API. The purpose of this study is to classify BISINDO letters A-Z and measure the accuracy, precision, recall, and cross-validation performance of the model. The dataset used was 4054 images with a size of consisting of 26 letter classes, which were taken by researchers by applying several research scenarios and limitations. The steps carried out are: dividing the ratio of the simulation dataset 80:20, and applying cross-validation (k-fold = 5). In this study, a real time testing using 2 scenarios was conducted, namely testing with bright light conditions of 500 lux and dim light of 50 lux with an average processing time of 30 frames per second (fps). With a simulation data set ratio of 80:20, 5 iterations were performed, the first iteration yielded a precision result of 0.758 and a recall result of 0.790, and the second iteration yielded a precision result of 0.635 and a recall result of 0.77, then obtained an accuracy score of 0.712, the third iteration provides a recall score of 0.746, the fourth iteration obtains a precision score of 0.713 and a recall score of 0.751, the fifth iteration gives a precision score of 0.742 for a fit score case and the recall score is 0.773. So, the overall average precision score is 0.712 and the overall average recall score is 0.747, indicating that the model built performs very well

Related SDGs Targets