|Abstract||This study uses image classification-based transfer learning to train models on the task of corrosion-detection on metallic surfaces. This is done by photographing images of samples of aluminium, iron and steel before and after corrosion to create visually differentiated datasets. With the exception of Model 1 which was trained and tested with a split of the original training set, the models were trained and tested on a newly prepared set to measure their accuracies fairly and realistically.
Model 1 was used to evaluate hypermeters, achieving an accuracy of 96.5%. Model 2, categorizing all images into corroded and uncorroded, scored an accuracy of 97.67%. Model 3, categorizing images into corroded, uncorroded and pitted, scored an accuracy of 95.67%. Model 4, trained to separate images into uncorroded aluminium, corroded aluminium, uncorroded steel/iron and corroded steel/iron, performed relatively poorly at 80%, but revealed that the majority of mislabeling is the result of combining the two materials in the sample model. Models 5 and 6 were trained on steel alone and aluminium alone, respectively. Model 5 scored an accuracy of 99.38%, while Model 6 scored a perfect 100%.