An Enhanced Multi-frequency Learned Image Compression Method

Learned image compression methods have represented the potential to outperform the traditional image compression methods in recent times. However, current learned image compression methods utilize the same spatial resolution for latent variables, which contains some redundancies. By representing different frequency latent variables with different spatial resolutions, the spatial redundancy is reduced, which improves the R-D performance. Based on the recently introduced generalized octave convolutions, which factorize latent variables into different frequency components, an enhanced multi-frequency learned image compression method is introduced. In this paper, we incorporate the channel attention module into multi-frequency learned image compression network to improve the performance of adaptive code word assignment. By using the attention module to capture the global correlation of latent variables, complex parts of the image such as textures and boundaries can be better reconstructed. Besides, an enhancement module on decoder side is utilized to generate gains. Our method shows the great visual appearance and achieves a better grade on the MS-SSIM distortion metrics at low bit rates than other standard codecs and learning-based image compression methods.