Deep learning–based automatic segmentation and 3d reconstruction in warhead performance testing

Warhead performance tests are of fundamental importance in the evaluation of shaped charge warheads (SCW) and fragmentation warheads (FW). These warheads are evaluated using the Depth of Penetration Test (DoPT) and the Arena Test (AT) \cite{lee2020deep}. While DoPT examines the penetration depth of metallic plates, AT analyzes the holes and defects formed on panels after detonation. In traditional methods, the manual inspection of test data is a labor-intensive, error-prone, and inconsistent procedure. Advances in deep learning and computer vision will enable the automation of this process, allowing highly accurate results to be derived from the test data. This thesis addresses two aspects of the data-processing workflow for warhead testing. First, two datasets were constructed using photographs obtained from two tests (DoPT and AT). The primary objective of both datasets is to distinguish, at the pixel level via semantic segmentation, the plates and the holes on their surfaces. However, due to the differing characteristics of the two datasets, different strategies were adopted. To enable pixel-wise hole detection, models that have demonstrated strong performance on datasets with similar characteristics, along with the associated loss functions, were reviewed in the literature. For both datasets, these models and loss functions were trained, and their performances were comparatively evaluated. In the experiments, MCCM-Net with Dice Loss achieved the best results in the SCW dataset with $\approx\ \%86$ IoU, while SegFormer with Cross Entropy achieved $\approx\ \%65$ IoU in the AT dataset. Following these results, annotation consistency was analyzed for both datasets, and differences in labeling on identical images were compared. Annotation agreement was found to be $80-85\%$ IoU for SCW and $49-63\%$ IoU for AT. These findings demonstrate that annotation consistency and architectural choices have a decisive effect on performance. The second part of the study primarily focuses on the innovative approaches that enhance detection performance. In this context, the Deep Supervision strategy strengthens gradient flow in intermediate layers, particularly improving the detection of small objects. OSuggested Learning Gate mechanism adaptively fuses encoder–decoder features at the pixel level to preserve local details and utilize contextual information more effectively. These contributions significantly increase segmentation accuracy while reducing the number of model parameters. Subsequently, the automatic framework incorporating pixel-level hole detections and the 3D reconstruction methodology are introduced. The proposed two-stage processing pipeline first separates metal panels from the background in low-resolution images, and then performs precise detection of holes and defects on high-resolution panel segments. After hole centers and contours are identified, they are converted into 3D coordinates using test data and re-projected onto the warhead geometry under both point-source and cylindrical surface assumptions. This enables the processing of a 4K (3840×2160) panel within seconds, providing a sustainable and repeatable workflow. In conclusion, the study presents an end-to-end approach extending from pixel-level segmentation to 3D fragmentation reconstruction, reducing manual analysis that could take tens of hours per panel to mere seconds. Methods such as Deep Supervision and Learning Gate are applicable not only to defense industry but also to domains such as medical imaging and industrial defect detection. Overall, this thesis introduces an innovative and comprehensive framework that integrates contemporary computer vision techniques with the practical needs of the defense industry.