eduzhai > Applied Sciences > Engineering >

Learning to Learn to Compress

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 6 pages

Abstract: In this paper we present an end-to-end meta-learned system for imagecompression. Traditional machine learning based approaches to image compressiontrain one or more neural network for generalization performance. However, atinference time, the encoder or the latent tensor output by the encoder can beoptimized for each test image. This optimization can be regarded as a form ofadaptation or benevolent overfitting to the input content. In order to reducethe gap between training and inference conditions, we propose a new trainingparadigm for learned image compression, which is based on meta-learning. In afirst phase, the neural networks are trained normally. In a second phase, theModel-Agnostic Meta-learning approach is adapted to the specific case of imagecompression, where the inner-loop performs latent tensor overfitting, and theouter loop updates both encoder and decoder neural networks based on theoverfitting performance. Furthermore, after meta-learning, we propose tooverfit and cluster the bias terms of the decoder on training image patches, sothat at inference time the optimal content-specific bias terms can be selectedat encoder-side. Finally, we propose a new probability model for losslesscompression, which combines concepts from both multi-scale and super-resolutionprobability model approaches. We show the benefits of all our proposed ideasvia carefully designed experiments.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×