Margin Based Siamese Network For Image Classification

Deep learning technology has become a popular research project in the computer field due to its rapid development in recent years.Thanks to the continuous rise of computing power,image classification technology has developed very rapidly with deep learning.The current research ideas for solving image classification mainly involve two types of frameworks: supervised learning and unsupervised learning.The discriminative model based on supervised learning is a relatively basic training framework.The breakthrough of this model is also one of the reasons for the rapid development of deep learning technology in the past 10 years.In recent years,more complex matric learning have become a new trend in the field of computer graphics.Siamese model is a case of matric learning.Since its introduction,Siamese model has been continuously improved for various computer image tasks.The research content of this article is to improve the Siamese network and make it suitable for the task of image classification.Due to the huge changes brought by the change of the angle of view and direction of the image itself,how to effectively identify images with many data types and small data volume is still a challenging task.A considerable number of research methods have tried to use the similarity of images to train the model,but have not noticed the effect of the interval variance of the image on the generalization performance of the model.This paper develops a more general neural network model Margin-Siamese for learning image similarity measures.By introducing the interval theory,the distribution of image pairs in the mapping space is expressed by margin,and the margin distribution is included in the index that measures the similarity of the image pairs in the mapping space.In this paper,Margin-Siamese is visualized on the MNIST data set,and sufficient experiments are performed on commonly used data sets.The results show that Margin-Siamese significantly improves the classification performance while reducing the generalization error.