Cascaded AI Task Decomposition: A Generalizable Framework for Multi-Stage Computer Vision Pipeline Validated on Automatic Number Plate Recognition for Nepali License Plates

Abstract

Detecting, localizing, and reading a license plate in a single pass usually lowers accuracy. Each of these sub-tasks is manageable on its own, but a single model must handle all of them together, across spatial scales that differ by more than an order of magnitude, so it settles for a compromise that is only moderate at each. This paper formalizes Cascaded AI Task Decomposition (CATD), which splits a compound vision problem into a sequence of tasks, each handled by a model trained only for that task and operating on a progressively cropped Region of Interest (ROI). The pipeline connects to any IP camera over the Real-Time Streaming Protocol (RTSP) and requires no hardware changes. To evaluate the framework, we built a three-stage Automatic Number Plate Recognition (ANPR) system for Nepali embossed plates: (1) a YOLOv8n-COCO vehicle detector, (2) a custom-trained YOLOv11n plate localizer, and (3) an RF-DETR Base character recognizer covering 61 classes. On the held-out test set the plate localizer reached 97.1 percent mAP@50 and the character recognizer 97.8 percent mAP@50. We integrated the pipeline into a parking-management system that automates the flow from plate capture to billing. We estimate the end-to-end precision at about 90.4 percent; this is not a direct measurement but a combination of the per-stage precisions, and validation on real Nepali embossed plates remains future work.