Computer Vision · 2026
SpectraScan
4th National — Mitsubishi Electric Cup 2026. 92.35% defect accuracy.
SpectraScan is an AI-powered defect detection system for industrial paint quality inspection, developed for the 6th Mitsubishi Electric Cup 2026 where it achieved 4th National Rank. The segmentation backbone combines DINOv2 feature extraction with an FPN-UNet decoder, reaching 92.35% detection accuracy and 86% dimensional validation precision on the competition dataset. Experiment tracking is handled by MLflow and hyperparameter search by Optuna, ensuring reproducible results across training runs.
Specifications
Features
DINOv2 + FPN-UNet segmentation architecture for pixel-level defect localisation
92.35% overall defect detection accuracy on paint surface imagery
86% dimensional validation precision for quality control measurement
MLflow experiment tracking for reproducible multi-run comparisons
Optuna automated hyperparameter optimisation for peak performance
4th National Rank — 6th Mitsubishi Electric Cup 2026