• 360° Panoramic Imagery
• Computer Vision
• Deep Learning
• Fiber Deployment Automation
• House Number Plate Detection
• Machine Learning
• Object Detection
• Text Recognition in Natural Scenes

Open Fiber aims to deliver the fastest internet services nationwide, ensuring that
every home can access high-speed fiber optic connectivity. Investing in cutting-edge
infrastructure and expanding network reach, the organization is committed to bridging
the digital divide and fostering innovation in urban and rural areas. However, due to a
lack of actual records, Open Fiber faces challenges in accurately estimating residential
buildings in specific regions, delaying infrastructure deployment timelines. Additionally,
relying on third parties for manual surveys is time-consuming and costly. At the same
time, digital records for small towns and rural areas are insufficient or inaccurate for
estimating the number of residential houses.
To address these challenges, we propose an AI/ML-based method that automates
the collection of house numbers from street view data, along with their corresponding
geo-coordinates. Our approach expedites the optical fiber infrastructure mapping
process, enhances deployment efficiency in rural and urban areas, manages the digital
record, and decreases overall costs. This technique can facilitate organizations and enable
them to make data-driven decisions for effective fiber optic infrastructure planning and
management.
The proposed methodology employs advanced deep learning models for object
detection and text recognition in a complex street view environment, including YOLOv8,
Yolov8-AM Faster R-CNN, SSD, and Detectron2, tailored for detecting house number
plates under diverse real-world conditions. A custom dataset is manually collected and
annotated to ensure comprehensive training coverage to achieve our objectives. For digit
recognition, we employ models such as ResNet, ViT+LSTM, Yolov8-AM DETR, fine-tuned
PaddleOCR, and EasyOCR, which are employed in a post-processing phase, enhancing
accuracy through integration with OCR frameworks.
The YOLOv8-AM model achieved top detection performance with 95% accuracy,
92% precision, and an mAP@0.5 of 0.926, surpassing YOLOv8, Faster R-CNN, and
Detectron2. For recognition, YOLOv8-AM also led with a precision of 91% and recall
of 93%, outperforming ViT+LSTM and ResNet.
The research also includes developing an application built on the deployed models.
This application is designed to visually present detected houses and their spatial
coordinates to estimate the distance from the nearest hub. This app integrates detection
and recognition outputs, providing a mapped visualization of identified house numbers
and their geographic locations. The application demonstrates how these models will
work in real world grounds, reducing manual processes and supporting precise, efficient
infrastructure expansion and management.