PRATAMA, VEMAS ADI (2026) IMPLEMENTASI MODEL COMPUTER VISION MULTI-OUTPUT UNTUK SISTEM BANTU TUNANETRA BERBASIS MOBILENETV2 DENGAN INTEGRASI SQUEEZE-AND-EXCITATION BLOCK DAN PENDEKATAN LATE FUSION. S1 thesis, Universitas Mercu Buana Jakarta.
|
Text (HAL COVER)
01 Cover.pdf Download (431kB) | Preview |
|
![[img]](https://repository.mercubuana.ac.id/style/images/fileicons/text.png) |
Text (BAB I)
02 Bab 1.pdf Restricted to Registered users only Download (37kB) |
|
|
Text (BAB II)
03 Bab 2.pdf Restricted to Registered users only Download (237kB) |
|
|
Text (BAB III)
04 Bab 3.pdf Restricted to Registered users only Download (65kB) |
|
|
Text (BAB IV)
05 Bab 4.pdf Restricted to Registered users only Download (277kB) |
|
|
Text (BAB V)
06 Bab 5.pdf Restricted to Registered users only Download (88kB) |
|
|
Text (DAFTAR PUSTAKA)
08 Daftar Pustaka.pdf Restricted to Registered users only Download (91kB) |
|
|
Text (LAMPIRAN)
09 Lampiran.pdf Restricted to Registered users only Download (151kB) |
Abstract
This research aims to implement, test, and analyze the performance gap of a multi output computer vision system based on MobileNetV2+SE using a Late Fusion approach to support independent mobility for the visually impaired on mobile devices. The model—which initially achieved validation accuracies of 97.45% (obstacle) and 95.33% (weather)—was converted to TensorFlow Lite format (26.0 MB) and integrated into a Flutter application. Testing was conducted on a mid-range Android device using two input schemes (static photo upload and real-time camera streaming) under bright and dim lighting conditions. The results showed the highest accuracy in the photo upload scheme under bright conditions, reaching 83% (obstacle) and 80% (weather). Switching to real-time streaming dropped the accuracy to 72%, while dim streaming conditions caused the most severe degradation, falling to 55% (obstacle) and 54% (weather) due to a majority class bias that triggered dominant misclassifications into the obstacle and foggy classes. While the system proves viable for mobile deployment under good lighting, it requires data augmentation incorporating lighting variations and motion blur simulations to improve its robustness in diverse real-world environments. Kata kunci: Visually Impaired, Computer Vision, MobileNetV2, Squeeze-and Excitation, Late Fusion, Mobile Deployment. Penelitian ini bertujuan mengimplementasikan, menguji, dan menganalisis performance gap sistem computer vision multi-output berbasis MobileNetV2+SE dengan pendekatan Late Fusion untuk mendukung mobilitas mandiri tunanetra pada perangkat mobile. Model dengan akurasi validasi awal 97,45% (obstacle) dan 95,33% (weather) dikonversi ke format TensorFlow Lite (26,0 MB), lalu diintegrasikan ke aplikasi Flutter. Pengujian dilakukan pada Android kelas menengah melalui dua skema input (upload foto statis dan streaming kamera real-time) dalam kondisi pencahayaan terang dan redup. Hasil menunjukkan akurasi tertinggi pada skema upload kondisi terang sebesar 83% (obstacle) dan 80% (weather). Peralihan ke streaming real-time menurunkan akurasi menjadi 72%, sementara kondisi streaming redup mengalami degradasi terbesar hingga 55% (obstacle) dan 54% (weather) akibat majority class bias yang memicu misklasifikasi dominan ke kelas obstacle dan foggy. Sistem ini layak untuk deployment mobile pada pencahayaan baik, namun membutuhkan augmentasi data variasi pencahayaan serta simulasi motion blur untuk meningkatkan robustness di lapangan. Kata kunci: Tunanetra, Computer Vision, MobileNetV2, Squeeze-and-Excitation, Late Fusion, Mobile Deployment
Actions (login required)
 |
View Item |
