Affiliations: [a] Faculty of Civil Aviation and Aeronautical, Kunming University of Science & Technology, Kunming, P.R. China | [b] Kunming SNLAB Tech Co., Ltd., Kunming, P.R. China | [c]
Material Certification Center, Ministry of Public Security, Beijing, P.R. China
Correspondence:
[*]
Corresponding author. Nan Pan, Faculty of Civil Aviation and Aeronautical, Kunming University of Science & Technology, Kunming 650500, P.R. China. E-mail: [email protected].
Abstract: The practical applications of the bullet rifling linear traces are severely restricted due of the complex shape and strong randomness. We propose a model of position and attitude parameters distribution at the end of specimens based on multimode elastic driving adaptive control method to achieve feature decomposition and error compensation correction of the attitude transformation of the specimen seat. The isolated forest algorithm was employed for abnormal processing of detection signals, non-small features were removed based on variable-scale morphological filtering algorithm, the trace curve profiles were extracted using the multiscale registration framework, and the square speed function optimization elastic shape metric algorithm was used to map the profiles into an embedding. Afterwards, a parametric shared conjoined triple deep learning model suitable for feature tracing and optimization of triplet selection and data augmentation strategies is proposed. This system is trained by minimizing a triplet loss function so that a similarity measure is defined by the L2 distance in this embedding. Finally, the trained model is used to do the similarity matching for the test set, try to solve the technical problems in the manual construction of guns and the inspection of bullet rifling traces.
