Transfer Learning Mask R-CNN for Radiograph Image Quality Indicator Localization
Merrill Dennison, Connor Seavers, Tsuchin Philip Chu · 2024
Ensuring the quality of industrial radiography is integral to its use for the inspection of components. This research work focused on applying an existing object detection and instance segmentation framework called Mask R-CNN to the recognition of image quality indicators (IQIs) in industrial radiographs. For the purposes of this project, a Mask R-CNN has been trained in a supervised manner, beginning with pre-trained ImageNet weights, toward the identification and localization of IQIs within digital radiographs. The goal of training the Mask R-CNN is for it to learn a mapping from the input radiographs to the output predictions of the bounding boxes and masks for each IQI in the input radiographs. On a high level, the Mask R-CNN serves as a function that takes a digital radiograph as input and provides a Python dictionary object as output. The output dictionary contains each region of interest predicted by the model for the given input radiograph, as well as their class IDs, probability scores, and mask images. Mask R-CNN is shown to be capable of adequately segmenting IQIs from radiographs when the standard practices for IQI placement are followed. This study explored the difference in Mask R-CNN performance when the training datasets are both small and contain IQIs at various orientations. A comparison is made between models trained with only parallel IQI examples and models trained with parallel, transverse, and askew IQI examples. This publication is focused on providing readers with a general understanding of the concepts, and numerical results are omitted in favor of visually depicting the best Mask R-CNN predictions. The results of this study have important implications for the application of existing computer vision and narrow artificial intelligence systems toward the detection of quality assurance objects within industrial radiography.