Generation of imagery-derived texture features for material discrimination

The Transportation Security Laboratory (TSL) conducts thorough assessments of explosives detection systems (EDSs), encompassing a wide range of explosive materials and hazardous substances. When necessary, inert simulants are employed, but they undergo a stringent verification process to accurately replicate specific properties of threat materials. Whether developed by the TSL or commercially acquired, simulants must undergo verification testing to ensure they mirror the desired threat properties. Historically, these assessments relied on rudimentary metrics like average density and effective atomic number, lacking insight into structural properties possibly being exploited by machine learning detection algorithms. Initial research focused on expanding the verification process by incorporating texture metrics extracted from computed tomography (CT) imagery aimed at deriving features that machine learning detection algorithms might also be utilizing. Two avenues of analysis were devised; first, we calculated 22 metrics through statistical analysis of pixel-based grayscale data, and second, we utilized a convolutional neural network (CNN) to classify images. Both of these methods were subsequently refined and are reported in this work. We augmented the number of metrics for the statistical analysis from 22 to 112, and within the CNN framework we harnessed the flattened array originating from the fully connected layer as a feature map. In both processes the analysis transitioned from a 2-dimensional to a 3-dimensional approach. We assessed the effectiveness of both procedures by testing them on imagery of 50 various materials, such as powders, liquids, putties, and emulsions, using Linear Discriminant Analysis (LDA) to evaluate their ability to distinguish between different materials. Finally, Principal Component Analysis (PCA) loadings were used to define 2-dimensional tolerance intervals for comparisons with loadings from other materials as a way to enhance the current simulant quality control process, ultimately improving the robustness of simulants.