evaluation

Code Architecture

This directory contains the core engineering and operational AI pipeline artifacts supporting the Edge AI document-detection system evaluated throughout the research project.

The codebase integrates computer vision training workflows, edge deployment orchestration, OCR-triggered processing, and downstream PHI/PII evaluation pipelines within a unified operational architecture.

---

Directory Structure

Directory	Purpose
core	Shared orchestration workflows supporting training, inference, and experimental execution
training	Training subsystem documentation and workflow architecture
inference	Real-time inference subsystem documentation
edge_pipeline	GStreamer, DeepStream, and edge runtime deployment workflows
evaluation	Evaluation subsystem documentation and metric-generation workflows
presidio	OCR and PHI/PII extraction pipelines using Microsoft Presidio
utilities	Supporting utility scripts and experimental helper workflows

---

Core Operational Workflow

Dataset Preparation
    ↓
YOLOv5 Training
    ↓
Cross-Validation Evaluation
    ↓
Edge Deployment
    ↓
Real-Time Inference
    ↓
Detection Stability Logic
    ↓
OCR Extraction
    ↓
Microsoft Presidio
    ↓
PHI/PII Triage Workflow

---

Primary Components

Shared Orchestration Workflow

core/doc_detector_yolov5_doccorner.py

Primary orchestration script supporting:

• dataset preparation
• YOLOv5 training
• five-fold cross-validation
• inference workflows
• metric generation
• visualization export
• optional corner regression
• optional OpenCV-based refinement
• GStreamer runtime execution

---

PHI/PII Processing Pipeline

presidio/pii_phi_pipeline.py

Primary OCR and PHI/PII extraction workflow supporting:

• OCR preprocessing
• Tesseract OCR execution
• Microsoft Presidio integration
• PHI/PII entity analysis
• downstream triage workflows
• JSON report generation
• document classification routing

---

Evaluation Utilities

evaluation/make_coco_gt_from_yolo.py

Utility workflow supporting conversion of YOLO-format labels into COCO-style ground-truth evaluation artifacts.

---

Experimental Utility Workflows

utilities/batch_pii_phi_eval_4class.py

Batch-processing utility supporting four-class PHI/PII evaluation workflows, confusion-matrix generation, and synthetic evaluation orchestration.

---

Deployment Environment

• NVIDIA Jetson AGX Orin 64GB
• Ubuntu 20.04
• CUDA
• TensorRT
• DeepStream 6.3
• OpenCV
• GStreamer
• YOLOv5
• PyTorch

---

Operational Focus Areas

• Real-time edge inference
• Computer vision deployment
• Privacy-preserving architectures
• Operational AI orchestration
• Embedded AI infrastructure
• OCR-triggered workflow processing
• PHI/PII triage integration

---

Research Context

The code architecture supports dissertation research evaluating whether localized Edge AI computer vision systems can function as upstream privacy-preserving control mechanisms capable of triggering downstream OCR and PHI/PII analysis workflows only after document detection events occur.

The repository emphasizes operational AI engineering, embedded computer vision deployment, and privacy-preserving Edge AI infrastructure within healthcare-oriented environments.