Digital Transformation of Visual Inspection Workflow Through Application of Computer Vision and Image Processing Techniques

Abstract The oil and gas industry relies heavily on visual inspection methods, which are non-destructive techniques that assess materials, components, or products to determine their current operational condition and develop appropriate maintenance schedules. These inspection processes can be performed manually without assistance or with the aid of specialized tools, providing critical initial overviews for both newly installed and in-service equipment across industrial facilities. The process involves detailed analysis of internal or external surface quality to identify and document superficial flaws that have developed during the equipment's operational lifetime. Currently, the majority of visual inspection workflows in the industry depend on experience-based rules and expert judgment, according to which trained inspectors assign damage severity levels to various types of equipment or components. The evolution and adoption of advanced image capture technologies, such as unmanned aerial vehicles (drones) and 360-degree cameras, have significantly reduced the requirements for direct physical inspections in hazardous or difficult-to-access field locations. However, scaling up assessment capabilities remains a highly labor-intensive effort and is expected to become a significant operational bottleneck as industrial sites progressively transition toward automated drone-based inspection methodologies. These modern imaging devices can capture enormous volumes of high-resolution images in a single inspection iteration, often exceeding 100,000 individual images per site visit. Therefore, the development and implementation of computer-assisted visual inspection workflows have become paramount for maintaining operational efficiency. This paper presents an approach to training and deploying sophisticated computer vision models specifically designed to accelerate the identification process and improve the accuracy of damage severity grading on industrial equipment. The proposed computer vision methodologies operate by learning and recognizing complex features associated with different categories of equipment damage. Multiple state-of-the-art computer vision model architectures, including ResNet-50, RegNet-z, and TinyViT, were extensively trained on detailed annotations of equipment geometries and damage polygon data. Model performance was rigorously compared using established metrics including pixel accuracy and intersection over union (IoU) measurements. IoU quantifies the overlap between predicted and ground truth regions, calculated as the ratio of the area of intersection to the area of union, with values ranging from 0 indicating no overlap to 1 representing perfect alignment. Prediction masks generated through patch inference were subsequently processed using conventional image processing algorithms, including morphological operations and contour approximation techniques, to effectively highlight and delineate prominent damaged regions. This work introduces a framework for developing computer vision applications with the specific aim of seamless integration into established visual inspection workflows of oil and gas facilities. The research highlights several significant improvements over existing literature in the field, including the standardization of annotation guidelines for consistent labeling, the application of lightweight hierarchical vision transformers for efficient damaged region identification, and the introduction of custom loss functions designed for robust and accurate damage classification.