AFWAN, Teuku M Arief and Gernowo, Rahmat and Wibawa, Helmie Arif (2026) MODEL PENGENALAN BAHASA ISYARAT INDONESIA (BISINDO) MENGGUNAKAN DEEP LEARNING LONG SHORT-TERM MEMORY UNTUK KOMUNIKASI INKLUSIF. Masters thesis, UNIVERSITAS DIPONEGORO.

	Text cover aja.pdf Download (56kB)
	Text COVER TESIS.pdf Restricted to Repository staff only Download (902kB)
	Text BAB I.pdf Download (77kB)
	Text BAB II.pdf Download (488kB)
	Text BAB III.pdf Restricted to Repository staff only Download (895kB)
	Text BAB IV.pdf Restricted to Repository staff only Download (847kB)
	Text BAB V.pdf Restricted to Repository staff only Download (58kB)
	Text DAFTAR PUSTAKA.pdf Download (126kB)
	Text LAMPIRAN.pdf Restricted to Repository staff only Download (2MB)

Abstract

Bahasa Isyarat Indonesia (BISINDO) merupakan sarana komunikasi utama bagi komunitas Tuli di Indonesia. Namun, ketersediaan dataset BISINDO yang representatif serta teknologi pengenalan otomatis yang akurat masih sangat terbatas. Penelitian ini bertujuan mengembangkan model pengenalan BISINDO berbasis deep learning menggunakan arsitektur Long Short-Term Memory (LSTM) yang dioptimalkan, serta membangun dataset video BISINDO yang terstruktur. Sebanyak 2.427 video dari 76 kelas dikumpulkan dan diproses menggunakan MediaPipe Holistic untuk mengekstraksi 1.662 fitur landmark per frame, yang kemudian dibentuk menjadi sekuens 30 frame. Model yang diusulkan berupa Enhanced LSTM dengan penambahan batch normalization, dropout, serta optimasi hyperparameter. Hasil pengujian menunjukkan akurasi pelatihan 95%, akurasi pengujian sebesar 93% dan confident score pada data stream langsung lebih dari 90%, serta performa yang stabil pada berbagai skenario gangguan. Penelitian ini menghasilkan dataset baru dan model pengenalan yang optimal, serta berkontribusi dalam pengembangan teknologi inklusif bagi komunitas Tuli di Indonesia dan membuka peluang penelitian lanjutan di bidang Sign Language Recognition (SLR).
Kata Kunci : BISINDO, Computer vision, MediaPipe holistic, Long short-term memory, SLR

Indonesian Sign Language (BISINDO) is the primary means of communication for the Deaf community in Indonesia. However, the availability of representative BISINDO datasets and accurate automatic recognition technologies remains limited. This study aims to develop a BISINDO recognition model based on deep learning using an optimized Long Short-Term Memory (LSTM) architecture, as well as to construct a structured BISINDO video dataset. A total of 2,427 videos from 76 classes were collected and processed using MediaPipe Holistic to extract 1,662 landmark features per frame, which were then organized into sequences of 30 frames. The proposed model is an Enhanced LSTM incorporating batch normalization, dropout, and hyperparameter optimization. The experimental results show a training accuracy of 95%, a testing accuracy of 93%, and a confidence score of over 90% in real-time data streams, demonstrating stable performance under various disturbance scenarios. This study produces a new dataset and an optimized recognition model, contributing to the development of inclusive technology for the Deaf community in Indonesia and opening opportunities for further research in the field of Sign Language Recognition (SLR).
Keywords: BISINDO, Computer vision, MediaPipe holistic, Long short-term memory, SLR

Actions (login required)

View Item