Deep Learning for Pipeline Interior Defect Classification: A Comparative Study of Polar and Cartesian Coordinate Representations

Recently, the development of deep learning has rapidly evolved, leading to increased demand for Artificial Intelligence (AI) driven assistance across various industries. This includes sectors such as oil and gas mining and sewer cleaning, where pipeline infrastructure maintenance often requires defect detection procedures. Within these industries, many studies have already explored the use of AI to enhance pipeline defect detection processes. Notably, some of these studies have employed a transformation technique that converts polar views of a pipeline interior into Cartesian views using polar-to-Cartesian coordinate principles, allowing for the analysis of unfolded pipeline images. After discovering this phenomenon, this work investigates whether transforming polar pipeline interior views into Cartesian views could increase model performance. Experiments involve training five renowned deep learning classification models with both original and unfolded images of pipelines. The experimental results indicate the original image performs slightly better as most of the models trained on the original dataset are better than the unfolded dataset.