mohanramasamy2005-eng/ai-pcba-board-detection

Purpose and entry points

This repository is an electronic-component/PCBA visual-inspection system built around YOLOv8 object detection. It supports interactive webcam inference, HTTP image prediction, model training, PostgreSQL detection logging, and a Streamlit monitoring view.

Primary launch points are:

• Webcam detector: app/main.py

• FastAPI service: app/api.py

• Training workflow: train.py

• Monitoring UI: dashboard.py

Model assets and inference boundary

YOLOv8 is the central ML runtime. The repository includes a PyTorch base/model artifact and an ONNX artifact:

• yolov8n.pt

• yolov8n.onnx

• models/.gitkeep

Inference-facing applications are expected to load an exported ONNX model (the README describes models/best.onnx as the trained output). ONNX forms the portability/efficient-inference boundary; PyTorch/Ultralytics is used for training and potentially local model handling.

Real-time inspection path

app/main.py is the local operational path for live quality inspection. It captures frames from a webcam through OpenCV, runs detector inference, and presents detections in real time. This path is intended for operator-facing manufacturing inspection rather than serving HTTP clients.

Relevant path:

• app/main.py

REST prediction service

app/api.py exposes the FastAPI application, with POST /predict accepting an uploaded image. The API boundary converts multipart image input into model inference results suitable for integration with external manufacturing or quality-control systems.

Relevant path:

• app/api.py

The service is run with Uvicorn (uvicorn app.api:app), binding application logic to an ASGI server/runtime.

Training and dataset contract

train.py drives YOLOv8 detector training. Dataset layout, split locations, class count, and labels are defined by:

• train.py

• data/data.yaml

The documented detection taxonomy has five classes: OK, Missing_Part, Scratch, Bent_Pin, and Soldering_Issue. Training outputs are expected to be exported for deployment, notably as ONNX artifacts consumed by inference paths.

Detection persistence and monitoring

Detection events are persisted to PostgreSQL through:

• log.py

The expected record contract is timestamp, detected class, and confidence score. dashboard.py is a Streamlit application that reads these stored detection records and presents a live-updating monitoring table:

• dashboard.py

• log.py

This creates a separate observability path: detector/API inference → PostgreSQL log records → Streamlit dashboard.

Deployment and infrastructure

The runtime is containerized by:

• Dockerfile

Cloud deployment guidance targets Google Cloud Run:

• gcp_deploy.md

Material external dependencies are Ultralytics YOLOv8, OpenCV, FastAPI/Uvicorn, PostgreSQL, Streamlit, Docker, and Google Cloud Run. Python dependency resolution is defined in requirements.txt.