All Projects
AI · Computer Vision · Deep Learning

Corrosion Annotator

Automated corrosion detection for industrial inspection — replacing manual visual assessment with a hybrid AI pipeline.

Overview

Hybrid segmentation system combining a pre-trained U-Net (ResNet34 backbone) with HSV classical computer vision, refined via CRF post-processing. Trained on a 3 GB multi-part annotated industrial dataset.

Key Features

  • Pixel-level corrosion segmentation — ensemble of deep learning + classical CV
  • 5-level severity classification (None / Low / Medium / Severe / Critical)
  • CRF boundary refinement for sharp mask edges
  • Applicable to pipelines, storage tanks, bridges, ship hulls, offshore platforms
  • Reproducible batch-processing workflow

Architecture

U-Net with ResNet34 backbone (ImageNet pre-trained) + HSV thresholding ensemble + CRF post-processing

What I Learned

The gap between a model that performs well on a validation set and one that works on real industrial photos is enormous. Lighting, rust color variance, image angle — the classical CV layer existed entirely to handle what the neural network couldn't generalise to.