Learned Bi-Resolution Image Coding using Generalized Octave Convolutions

Learned image compression has recently shown the potential to outperform standard codecs. State-of-the-art rate-distortion (R-D) performance been achieved by context-adaptive entropy coding approaches in which hyperprior and autoregressive models are jointly utilized effectively capture spatial dependencies latent representations. However, latents feature maps of same resolution previous works, contain some redundancies that affect R-D performance. In this paper, we propose a learned bi-resolution approach is based on developed octave convolutions factorize into high low components. Therefore, redundancy reduced, improves Novel generalized convolution transposed-convolution architectures with internal activation layers also proposed preserve more structure information. Experimental results show scheme outperforms all existing methods as well codecs such next-generation video VVC (4:2:0) both PSNR MS-SSIM. We can improve other auto-encoder-based schemes semantic segmentation denoising.