Computer VisionDINOv2OpenVINOIndustrial AIMitsubishi Electric Cup

SPECTRASCANN: BUILDING A REAL-TIME INDUSTRIAL AI PAINT DEFECT DETECTION SYSTEM

SpectraScan (shipped as CON-SOL-E Vision Pro) is a full-stack industrial AI platform for automotive paint quality inspection. It won 4th National Rank at the 6th Mitsubishi Electric Cup 2026 competing against teams from IITs, NITs, and top engineering colleges.

4th

National Rank

92.35%

Segmentation accuracy

<500ms

End-to-end latency

3

Defect classes detected

THE PROBLEM: PAINT DEFECTS IN AUTOMOTIVE MANUFACTURING

Automotive door panels reject rates from paint defects — dust inclusions, scratches, and rundown (drips) — cost manufacturers millions annually. Manual visual inspection is inconsistent, subjective, and slow. The brief: build a system that detects all three defect classes in real-time and triggers a PLC-controlled CNC gantry for automated remediation.

ARCHITECTURE OVERVIEW

Camera capture (FLIR industrial) ↓ DINOv2 ViT-B/14 backbone (OpenVINO INT8, ~180ms) ↓ FPN-UNet segmentation head (defect mask generation) ↓ FastAPI SSE stream → Electron.js desktop client ↓ PLC serial bridge → CNC gantry XY positioning ↓ PDF inspection report (defect coordinates, severity)

WHY DINOV2 + OPENVINO?

DINOv2's self-supervised ViT features generalise remarkably well to industrial textures without domain-specific pretraining — critical when you have limited labelled defect data. OpenVINO INT8 quantisation on an Intel iGPU brought inference from 850ms to sub-500ms without meaningful accuracy degradation, enabling real-time feedback on the production line.

Alternatives considered: YOLOv8-seg (too fast but lower precision on subtle paint texture defects), SAM (too slow for real-time), ResNet50 FPN (insufficient feature richness for the rundown class).

THE RAG CHATBOT MODULE

SpectraScan ships an embedded local RAG chatbot (Phi-3-mini + ChromaDB) that allows factory operators to query inspection history, defect statistics, and maintenance logs in natural language — no cloud dependency, all inference runs on-device. This was built using LangChain with a custom chunking strategy optimised for structured PDF inspection report text.

TRAINING DETAILS

Dataset: Custom-labelled automotive door panel images 3 classes: Dust, Scratch, Rundown Backbone: DINOv2 ViT-B/14 (frozen) Head: FPN-UNet (trained from scratch) Tracking: MLflow experiment tracking Tuning: Optuna hyperparameter search (150 trials) Accuracy: 92.35% mIoU on test split
View on GitHub →