基于多任务级联残差网络的枪支图像识别系统

针对枪支种属识别目前依赖检验人员经验、识别效率较低的问题，建立一种基于多任务级联深度残差网络的枪支图像自动识别模型。以ResNet18为基本构建单元，通过级联融合4个任务中的Softmax损失函数约束，实现对枪支图像从枪族到枪型的多维度聚类。在该模型的基础上，设计一套制式枪支图像智能检索系统，对拍摄上传的枪支图像种属信息进行自动识别。在自建的制式枪支图像数据集上进行实验，结果表明，与EfficientNet、NTS-net等模型相比，该模型的识别准确率更高，Rank-1、Rank-20识别准确率分别达到61.12%、95.28%，且其具有更好的鲁棒性。

Firearm Image Recognition System Based on Multi-Task Cascaded Residual Network

The traditional firearm recognition methods rely heavily on expertise, and are limited in recognition accuracy.To address the problem, a model for automatic firearm image recognition is built using multi-task cascaded deep residual network.With ResNet18 as the basic build block, this model fuses the Softmax loss function constraints in the four cascaded tasks and realizes firearm image clustering, which is based on multiple dimensions ranging from the firearm family to the specific gun type.Based on the proposed model, a system for intelligent firearm image retrieval is designed, which can automatically recognize the type of the firearms in uploaded images.The experimental results on a self-made firearm image dataset show that the model displays a higher recognition accuracy in Rank-1(61.12%) and Rank-20(95.28%) than EfficientNet, NTS-net and other models.The proposed model also provides better robustness for gun image recognition in real scenes.