Learning Visual Representations With Deep Neural Networks

Recently,with the development of technology,more and more visual data can be obtained from the internet.Learning useful information from the large quantity of data has become a challenging task.Deep learning methods have played an important role in processing the big visual data.Deep neural networks are a kind of machine learning models that can learn hierarchical representations from raw data,which will be beneficial to the subsequent visual tasks such as image recognition,scene image segmentation.However,deep networks usually require a large number of samples to train and are prone to overfitting.In this paper,we study these problems and provide some possible solutions to improve the networks’performance.The contributions of our paper are as follows:1)we propose a novel kind of deep unsupervised network,named the stacked convolutional denoising auto-encoders,to learn deep representations from the raw images without any label information.2)we propose a novel kind of regularizer,called structured decorrelation constraint,to prevent the network from overfitting and improve the networks’ generalization.