MIXLIC: Mixing Global and Local Context Model for learned Image Compression
Haihang Ruan, Feng Ryan Wang, Tongda Xu, Zhiyong Tan, Yan Wang · 2023
Learned Image Compression (LIC) is considered as a future direction for image compression. However, existing LIC methods only marginally outperform the latest traditional codec VVC. This is because current LIC methods have not fully utilized the global and local information of images. In this paper, we propose a MIXing global and local context Module (MIXM), which combines global context extractor (GCE) with local context extractor (LCE) in a parallel design, capturing both global and local dependencies. Based on the MIXM, we further build the MIXLIC, a mixing global and local context model for learned image compression, which can fully utilize global and local context to improve performance. Experimental results show that our proposed method MIXLIC achieves state-of-the-art RD performance and gets more visually pleasant results compared with other learning-based methods and traditional codecs.