Frequency-Aware Learned Image Compression Using Channel-Wise Attention and Restormer

In recent years, learned image compression has leveraged convolutional neural networks (CNNs) and achieved notable advancements in coding efficiency. However, learned image compression has a limitation of balancing global context and local texture because the global structure easily ignores local redundancy, especially for the non-repetitive textures, affecting the reconstruction performance. In this paper, we propose a frequency-aware learned image compression network based on channel-wise attention and Restormer. In encoder, we perform locality-aware encoding based on window-based attention. In decoder, we combine Restormer for global context and channel attention for local texture, forming a mixed attention module. Moreover, we design a frequency-aware network based on the discrete wavelet transform (DWT) in a hyperprior network to learn both spatial and frequency features, while removing distortion and capturing significant context in the low frequency domain. Considering the training instability and convergence difficulty of Restormer, we propose a novel training strategy based on knowledge transfer to optimize compression models during training. Experimental results demonstrate that the proposed method reconstructs high quality images with outstanding coding efficiency and achieves average BD-rate gains of 9.88% and 3.24% in terms of PSNR and MS-SSIM on the Kodak24 dataset over the auto-regressive hyperprior model proposed by Cheng et al., respectively.