Image Compression Architecture with Built-in Lightweight Model

Many studies have applied deep learning techniques to image compression in recent years. However, it is difficult to apply these algorithms to all sorts of images because it is necessary to cover a large number of diverse images for training. To tackle this challenge, we propose a new image compression framework based on customized learning. We only let the model analyze a single image and learn the lost information after traditional compression algorithms. Then, we encode the trained model parameters and the compressed image separately and transmit them together. At the decoder, we can restore the uncompressed image content by the model parameters. In our experiments, we use JPEG and JPEG2000 to validate our algorithm, and from the experimental results we can prove that our framework is feasible.