Synthetic generated data for intelligent corrosion classification in oil and gas pipelines

This research presents the K-Pipelines dataset, a pioneering synthetic image collection designed specifically for the classification of corrosion in oil and gas pipelines. Instead of training custom generative architectures, our research used an online image generation tool powered by Stable Diffusion. This choice leveraged the platform’s robust capability to quickly produce a high volume of diverse and detailed images, saving significant time and resources. The dataset was carefully constructed using a sequence of refined prompts, derived from a review of pipeline characteristics including material types, environments, and corrosion forms. K-Pipelines consist of 600 PNG images of 512 × 512 resolution. Furthermore, an augmented version was developed, totaling 1080 images. Our evaluation employed state-of-the-art deep learning classifiers, specifically VGG16, ResNet50, EfficientNet, InceptionV3, MobileNetV2, and ConvNeXt-base, to test the integrity of the K-pipelines dataset. These models showcased its robustness by consistently achieving accuracies around the 90% mark, illustrating the dataset’s substantial promise as a resource for both AI research and real-world applications in the oil and gas industry. The dataset is publicly available for access and use within the scientific community. • A new dataset for corrosion classification in pipelines is introduced. • The images are generated with online synthetic generation tools. • The images include different types of corrosion, material, and environment. • Two versions available, one with 600 images, and another augmented with 1080 images. • Computer vision architectures were tested, reaching accuracies above 84%.