An Image Recognition And Application Platform Based On Barycenter Learning

In recent years,with the continuous upgrading of hardware performance,more and more artificial intelligence applications based on deep learning are constantly entering our lives.The applications all provide great convenience,from voice assistants and machine translation which are based on natural language processing technology,to image recognition,detection and segmentation which are based on computer vision technology.However,it is worth noting that the success of deep learning applications depends on the scale of the training data.For example,the model trained on the large-scale image dataset Image Net [1] can achieve good generalization capability in various data scenarios.Otherwise,the model trained on a small-scale dataset will be too limited caused by over-fitting problem.Therefore,this article focuses on data augmentation techniques in deep learning to expand small-scale datasets to enhance the generalization ability of the model.We mainly research on image data.It can be simply implemented by adding a small disturbance to the original image or changing the local information of the image.These methods have been proved to be effective.Recently,a number of researches on image mixing have emerged and have also achieved good performance.After our research and discovery,we can re-think these image mixing methods as the process of finding image barycenters.For instance,the Mixup [2] method is essentially finding the Euclidean barycenter.Therefore,based on this view,we propose two novel methods to generate image barycenters.One is the Opt Trans Mix method based on Wasserstein barycenter,and the other is the Auto Mix method based on latent barycenter in the hidden space of the deep neural network.The effectiveness of the two methods in data augmentation has been verified through comparative experiments.They also achieved good performance in terms of openset problem and robustness against noise.In addition,we also set up an image recognition platform to quickly verify the computer-vision-related requirements.We integrate the two image mixing methods proposed in this paper into the platform as an effective way of data augmentation for small-scale dataset.During the master's degree,the platform has been effectively applied in the evaluation process of various projects.