Image Compression with Deeper Learned Transformer

Deep learning is known for its flexibility and infinite potential to approximate any function. Is it possible to approximate image compression using deep learning? The answer is yes. This article compares three major deep learning techniques used in image compression now and proposed an approach with deeper learned transformer and improved optimization goal, which achieved improved peak signal-to-noise ratio (PSNR) and multi-scale structural similarity (MS-SSIM) under very low bits per pixel (bpp). Experimental results show that the proposed approach outperformed BPG (RGB 4:4:4) in natural scene images compression, and is capable to handle arbitrary image shapes, which makes it applicable to practical image compression workloads.