PCB Defect Detection (YOLOv8 Fine-Tuning)

This repository contains the code and analysis material for a PCB surface-defect detection study based on YOLOv8.
The project compares fine-tuning strategies for YOLOv8m and YOLOv8m-P2, evaluates detection quality on held-out data, and includes post-training confidence-threshold analysis for operating-point selection.

Scope of the Experiments

The study focuses on four model/strategy variants:

• YOLOv8m-P2 (full fine-tuning)
• YOLOv8m (full fine-tuning)
• YOLOv8m (backbone frozen; neck + head trainable)
• YOLOv8m (head-only fine-tuning)

Repository Layout

• main.py: main pipeline entry point (training, evaluation, prediction)
• configs/: experiment configurations
• scripts/: helper scripts for experiment execution
• utils/: data handling, training utilities, evaluation, logging, plots
• EDA/: dataset exploratory analysis notebook
• analyses/: result analysis, architecture inspection, and visualization notebooks
• artifacts/: locally stored run outputs used for offline analysis