2018/03/01 by Masanori Suganuma, Suganuma, Masanori, Mete Özay +3 · 1 citation
Computer Science · Engineering · #Advanced Image Processing Techniques #Image Processing Techniques and Applications #Digital Media Forensic Detection
paper · pdf · doi:10.48550/arxiv.1803.00370
Researchers have applied deep neural networks to image restoration tasks, in\nwhich they proposed various network architectures, loss functions, and training\nmethods. In particular, adversarial training, which is employed in recent\nstudies, seems to be a key ingredient to success. In this paper, we show that\nsimple convolutional autoencoders (CAEs) built upon only standard network\ncomponents, i.e., convolutional layers and skip connections, can outperform the\nstate-of-the-art methods which employ adversarial training and sophisticated\nloss functions. The secret is to employ an evolutionary algorithm to\nautomatically search for good architectures. Training optimized CAEs by\nminimizing the \ℓ2 loss between reconstructed images and their ground\ntruths using the ADAM optimizer is all we need. Our experimental results show\nthat this approach achieves 27.8 dB peak signal to noise ratio (PSNR) on the\nCelebA dataset and 40.4 dB on the SVHN dataset, compared to 22.8 dB and 33.0 dB\nprovided by the former state-of-the-art methods, respectively.\n