Block based learned image compression

Efficient image compression is very important for storage, retrieval, processing and transmission of image contents. The objective is to find a striking balance between compression ratio and the distortion in image. Recently, there has been a rise in interest on lossy neural network based compression algorithms. Specifically, autoencoder based compression schemes have shown great potential in learned image compression domain. This paper proposes a new algorithm for learned image compression using block based Generative Adversarial Networks. The adversarial network was trained on the blocks derived from a large image data-set. The compressed images were compared against standard compression schemes such as JPEG, PNG to show the comparative strength of block based learned compression algorithms. It has been found that performance of algorithm drops significantly at low bits per pixel. So, the paper compares the algorithm performance at various bpp values.