Concurrent in-situ flash X-ray radiography, Photon Doppler Velocimetry (pdv), and high-speed video using the ARL HIDRA provides microsecond-resolved information during ballistic impact experiments. Manual image analysis, although laborious and time intensive, provides characterization of dwell time, penetration velocity, and timing of break-out of the back surface of the target. To speed up the process and provide reproducible results image segmentation workflows were created using scripted image processing and machine learning in the MEDE Data Science Cloud (MEDE-DSC).
Scripted Image Processing (SIP) of HIDRA radiographs includes pre-processing to normalize contrast and perform film registration. A convolutional neural network with three hidden layers and pixel-level class masks (Mask R-CNN) was built and pretrained on the MS-COCO dataset. Manually landmarked experimental images were utilized for final training. K-fold cross validation provided unbiased training set selection. Precision-recall evaluation metrics were implemented for model comparison and hyperparameter tuning. The resulting model provides accurate masks with impressive timing. Data reduction has been reduced from approximately 8 hours per experiment to under a second