<![CDATA[Deaf and speech-impaired individuals in Indonesia face communication barriers due to limited public understanding of sign language. In real use, SIBI communication often occurs in dim lighting, yet recognition models are mainly evaluated under normal illumination, motivating robust low-light recognition. This study develops a CNN model based on MobileNetV2 to recognize SIBI (Indonesian Sign Language System) letter gestures under low-light conditions (50-100 lux). The dataset comprises 5,579 images of 26 SIBI letters, divided stratified 80:10:10. The methodology includes preprocessing with Bilateral Filter, CLAHE in LAB color space, and Adaptive Gamma Correction, plus transfer learning and fine-tuning with data augmentation. Evaluation results show 97.13% test accuracy, with most errors among similar letters. Real- time testing is stable within 50-100 lux, though accuracy decreases below 50 lux or with shadows. These findings indicate that the proposed preprocessing methods and MobileNetV2 CNN maintain reliable SIBI recognition in low-light environments.]]>
Copyrights © 2026
