This study tested a machine-learning model meant to provide objective support for a physical fit examination of duct tapes.
This paper describes the construction and use of a machine-learning model to provide objective support for a physical fit examination of duct tapes. This pilot study demonstrates the feasibility of computational algorithms to build physical fit databases and automated comparisons using deep neural networks, which can be used as a model for other materials. The authors present the ForensicFit package that can preprocess and database raw tape images. Using the processed tape image, the authors trained a convolutional neural network to compare tape edges and predict membership scores (i.e., fit or non-fit category). A dataset of nearly 2000 tapes and 4000 images was evaluated, including various quality grades: low, medium, and high, as well as two separation methods, scissor-cut and hand-torn. The model predicts medium-quality and high-quality scissor-cut tape more accurately than hand-torn, whereas for low-quality tape predicts the hand-torn tapes more accurately. These results are consistent with previous studies performed on the same datasets by analyst examinations. A method of pixel importance was also implemented to show which pixels are used to make the decision. This method can confirm some fit features that correspond with analyst-identified features, like edge morphology and backing pattern. (Published Abstract Provided)