Feature Recognition of Cartridge Case Traces Based on Improved Attention Mechanism and NetVLAD

Aiming at the traditional cartridge case trace recognition methods mostly stay in the manual recognition stage, which makes the accuracy of cartridge case trace recognition has great defects, we propose a cartridge case trace feature recognition method based on the improved NetVLAD network. Firstly, we collect images of cartridge case traces of seven types of establish guns and an image database, with NetVLAD network as the base model and ResNet-18 as the backbone network. Secondly, we introduce the improved attention mechanism CBAM (Convolutional block attention module). Finally, for the convolutional kernel scale single problem, we add one-dimensional convolution and multi-scale cavity convolution module to improve it, fuse different sensory field features, in order to reduce the dimensionality and also effectively inhibit the feature loss. The final network model generates more accurate and compact VLAD (Vector of Locally Aggregated Descriptors), which are used for feature matching between trace images, thus improving the image retrieval accuracy. The experimental results show that the proposed model has a better recognition and matching ability on the cartridge case trace dataset of 7 types of guns, with an accuracy of more than 95%, and compared with other neural networks, the proposed model compares the accuracy of the NetVLAD model with VGG-16, ResNet-50, ResNet-18, AlexNet, and InceptionV3 as the backbone network by 19.07%, 17.07%, 17.07%, and 17.07%, respectively, improved by 19.07%, 17.16%, 14.58%, 16.51% and 21.47%. Moreover, the matching accuracy of the proposed model is improved compared with the pre-improved model on the cartridge traces dataset of seven types of guns, which can provide a reference for the classification of gun types in the field of cartridge traces identification.