Learned image compression with generalized octave convolution and cross-resolution parameter estimation

• Learned Image Compression Generalized Octave Convolution Outperforms the state-of-the-art deep learning-based methods and traditional codecs such as VVC Recently, image compression approaches based on learning have gradually outperformed existing standards including BPG intra coding. In particular, application of context-adaptive entropy model significantly improves rate-distortion (R-D) performance, in which hyperpriors autoregressive models are jointly utilized to effectively capture spatial redundancy latent representations. However, representations still contain some correlations. addition, these cannot be accelerated decoding process by parallel computing devices, e.g. FPGA or GPU. To alleviate limitations, we propose a learned multi-resolution framework, exploits recently developed octave convolutions factorize into high-resolution (HR) low-resolution (LR) parts, similar wavelet transform, further R-D performance. speed up decoding, our scheme does not use model. Instead, exploit an additional hyper layer encoder decoder remove representation. Moreover, cross-resolution parameter estimation (CRPE) is introduced proposed framework enhance flow information improve An information-fidelity loss total function adjust contribution LR part final bit stream. Experimental results show that method separately reduces time approximately 73.35 % 93.44 compared with methods, performance better than H.266/VVC(4:2:0) both PSNR MS-SSIM metrics across wide rates.