RustSEG

RustSEG is a corrosion segmenter that was developed for my final year engineering honours project at the Australian National University. It was inspired from and heavily influenced by corrosiondetector.com. The segmenter uses unsupervised learning in that it never receives any per-pixel training data. The system works by passing through a bespoke classifier built on U-Nets encoder side. If this detects corrosion it is passed to a Grad-CAM++ module that determines which pixels helped make that decision. This method returns a 28x28 heatmap, then using post processing the system can calculate a mask for rust in the image.