Compressão de imagens com perda utilizando redes neurais artificiais

Abstract

With the rapid growth of media currently being transmitted over the Internet, lossy image compression algorithms are indispensable. Therefore, there are several studies that seek to improve the compression algorithms, in order to obtain higher image quality and higher compression rates. The objective of this work is to present an image compression algorithm using Artifcial Neural Networks based on the autoencoder network architecture. The models created, when compared to the JPEG algorithm and applied in a set of binary images, reached compression rates up to 7 times higher with a structural similarity index (SSIM) 19 % higher. When applied to a set of grayscale images, achieved compression rates up to 21 times higher, and SSIM 6 % below JPEG. And when applied to a set of colored images achieved a compression rate up to 2 times higher, with an SSIM 6 % above JPEG.