Feature Enhanced Learning Image Compression With Recurrent Criss-Cross Attention

Learning image compression has aroused active exploration and extensive research in recent years. However, the existing compression models fail to capture the complex texture representation of images adequately, causing the loss of detailed texture in reconstructed images. In this paper, we propose a feature enhanced learning image compression method. First, we introduce a novel recurrent criss-cross attention (RCCA), which could effectively extract the global features and contextual information of the input image. Next, we combine all the generalized divisive normalization (GDN) of the encoder to construct a pyramid feature fusion (PFF) module. Then, we introduce an enhanced inverted bottleneck residual blocks (Enh RBs) at the decoder to expand the high-dimensional channel space of texture detail representation. Finally, the experimental results show that our method can produce superior results compared to the state-of-the-art methods.