PROJECT TITLE :

Learning a Single Tucker Decomposition Network for Lossy Image Compression With Multiple Bits-per-Pixel Rates

ABSTRACT:

A common problem in Image Processing is lossy image compression (LIC), which seeks to use inexact approximations to reduce the size of an image. An encoder-quantizer-decoder network can be learned using deep convolutional Neural Networks (CNNs) from a vast amount of data. CNN-based LIC approaches, on the other hand, can only train a network to handle a particular number of pixels per bit-per-pixel (bpp). In practical LIC applications, a "one network per bpp" problem restricts CNNs' generality and flexibility. LIC at several bpp rates can be learned from a single CNN in this work. Decomposing a latent picture representation into a series of projection matrixes and an essential core tensor is accomplished using a novel Tucker Decomposition Layer (TDL). Latent image representation in CNNs can be easily manipulated by modifying the core tensor rank and its quantization. A coarse-to-fine training technique is also introduced to recreate the decompressed pictures, using an iterative non-uniform quantization scheme. TDNet's PSNR and MS-SSIM indices have been extensively tested to demonstrate its state-of-the-art compression capability.