Edge-enhanced Generative Adversarial Network for Reconstruction of Compressed Image

The images are often compressed to reduce storage usage or accelerate image transmission. However, the compression process always results in the loss of image details, such as edge details, which degrades the visual experience. Plenty of reconstruction methods have been proposed, but it is yet challenging to enhance edge details more precisely. In this paper, we propose a GAN-based image reconstruction architecture mainly for edge enhancement. Our model improves by cycleGAN; the model's input adds the extracted edge of the compressed image to promote generating more precise edge information. To further optimize the image edge details, we define a new edge loss function to improve the quality of the generated image. Lastly, we train and test the images from the CelebA dataset and the ACDC medical dataset. The experimental results show that the reconstructed images are clear under the high compression ratio and have more precise image edge details.