Improved Generative Adversarial Networks For Better Compression Quality And Structural Similarity

Authors

  • S.Aruna Deepthi, E.Sreenivasa Rao

Abstract

Traditional image compression algorithms hamper the original quality of the images introducing artifacts and quality degradation. In this paper, we propose the idea of compression using generative models  more precisely and visually accurate regeneration is done at a better compression rate for both video and image data. We design GAN framework comprising an encoder, generator and a discriminator that is trained combined to realise the objective of generative learned compression.  In a conventional GAN, the encoder-decoder based on deep neural networks trained with  L1 and L2 loss can miss the details, appear blurry and unrealistic while reconstructed from the compression format.  We have resolved this issue with a novel GAN technique that addresses the various image artifacts resulting from compression by selecting a multi-variate loss function that reduces both the structural and adversarial loss. For the deep neural network design for compressor and decompressor units, we have used a tailor-made architectural approach that involves four convolutional layers based on residual networks (ResNet), while for the discriminator/classifier that identifies the fake or real pair, we have adapted a conventional approach that comprises of convolution and dropout layers. On successful implementation of the proposed architecture and loss function, we have achieved varied compression rates at different levels of network training and produce reconstructed images similar to the original images with better structural similarity.

Keywords- GAN, JPEG compression, SSIM,DL,NN

Published

2020-12-31

Issue

Section

Articles