Unsupervised Feature Learning With Symmetrically Connected Convolutional Denoising Auto-encoders

Convolutional neural networks (CNNs) have shown their power on many computer vision tasks.However, there are still some limitations, including their dependency to large scale labeled training data and sensitivity to weight initialization.In this paper, we try to address these two problems by proposing a simple yet powerful CNN based denoising auto-encoder which can be trained end-to-end in an unsupervised manner.The network architecture we employ is a fully convolutional auto-encoder with symmetric encoder-decoder connections.The proposed method can not only reconstruct clean images from corrupted ones, but also learn image representation through the reconstruction training.It can further be adapted to find data driven network initialization without using extra training data.Experimental results show that our network can learn good feature from unlabeled data, which can be easily transferred to different high level vision tasks such as image classification and semantic segmentation.The data driven initialization method based on the convolutional auto-encoder is also competitive.