Siamese Convolutional Neural Network For Genuine And Fake Coin Image Classification

Coins are widely used in everyday life and apart from being used as currency,some people also like to collect coins.However,in recent years,a large number of illegal counterfeiting rings have been manufacturing and selling fake coins,causing huge losses and damage to society.Although there are forensic experts to examine suspected fake coins,this is unrealistic considering the huge number of coins that need to be examined.Therefore,the use of modern high-tech means to identify genuine and fake coins is crucial for fake detection,e.g.deep learning-based methods that not only improve the accuracy of coin detection,but also save a lot of manpower as well as material resources.This paper focuses on siamese convolutional neural network-based image classification of genuine and fake coins.In this paper,an end-to-end siamese neural network is proposed to extract features of genuine and fake coin images and classify them based on the extracted features.This network consists of two branches that share weights between them.Each branch in the network extracts features of the image using a convolutional neural network,and an attention mechanism is added behind the convolutional layer of each backbone network for enhancing feature representation in different directions.A hierarchical bilinear pooling module is also introduced,which not only supports layer-to-layer interaction,but is also used to learn fine-grained representations in a form that enhances each other.In addition,two loss functions are used in the network structure proposed in this paper,the cross-entropy loss function and the contrastive loss function.The contrastive loss function compares the difference between the actual output and the expected output,and ensures that the difference between image features is as small as possible when the input is the same class of image;conversely,the difference between image features is as large as possible.The two sub-branches of the system are each classification models,and the loss functions are both cross-entropy loss functions.The experimental results show that the performance of the network with both loss functions is higher than that with only one loss function,which shows that the two loss functions play a good mutual reinforcement role with each other.To demonstrate the effectiveness of the method proposed in this paper,a series of ablation experiments and comparison tests were conducted.The effectiveness of the attention mechanism,the cross-layer bilinear pooling module and the Siamese network is demonstrated using the ablation experiments.Overall comparative experiments were conducted in five aspects,including a traditional manual feature-based approach,a Vi T model based on the Transformer idea,a generic convolutional neural network-based approach,a fine-grained classification-based approach and a coin study-based approach.The experimental results show that the method proposed in this paper achieves 98.6% and 98.4% accuracy in the datasets DanishCoin1991 and DanishCoin2008 respectively,which is higher than the existing methods.