A Verification Method of Industrial Metal Parts using Siamese Residual Network

Effective verification methods have a pivotal role in curbing the counterfeiting of industrial metal parts. This paper proposes a verification method by extracting the surface textures from a single image of industrial metal parts through a Siamese residual network. The proposed method expresses texture features as a 32-dimensional feature vector. The L2 distance between two feature vectors represents their similarity, further indicating the authenticity of industrial metal parts. Feature fusion architecture in the proposed network is beneficial to utilize surface textures from multiple scales, which ensures the completeness of features. Multiple loss functions, namely triplet loss and cross-entropy loss, are applied in the training phase. The combination of multiple loss functions improves the accuracy of the network while accelerating the training process. The effectiveness of the proposed network is proved by experiments and the performance is evaluated. The network achieves 97.31±0.52% accuracy on the test dataset, enabling reliable anti-counterfeiting verification of industrial metal parts.