Abstract: A point cloud deep learning method is presented to identify the pitting corrosion damage in steel structures for accurate three-dimensional quantification of damage features. Point cloud data of corroded steel surfaces were acquired through salt spray accelerated corrosion tests and industrial-grade three-dimensional scanning. A deep learning dataset was established with a pitting corrosion semantic annotation algorithm, and the generated data were used for dataset augmentation. A point cloud deep learning architecture was used to learn pitting corrosion semantic features, and point cloud convolution modules were introduced to strengthen network learning. The results show that the point cloud deep learning architecture accurately identifies pit features with an accuracy exceeding 81% and a processing efficiency much higher than that of classical algorithms. The convolution modules markedly improve the recognition of pitting corrosion semantic features, and their optimization effect on model training is most evident when the proportion of generated data is about 15% of the original dataset. Compared with the classical corrosion pit extraction algorithm based on multi-height-threshold slicing and connected component analysis, the deep learning model significantly improves the processing efficiency, and the minimum computation time is only 0.06% of that of the classical algorithm.