Deep Networks for Image-to-Image Translation with Mux and Demux Layers

Hanwen Liu, Pablo Navarrete Michelini, Dan Zhu; Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0

Imageprocessingmethodsusingdeepconvolutionalnetworks have achieved great successes on quantitative and qualitative assessments in many tasks, such as super–resolution, style transfer and and enhancement. Most of these solutions use many layers, many filters and complex architectures. It is difficult to implement them on mobile devices, e.g. smart phones, because of the limited resources. Many applications need to deploy these methods on mobile devices. But it is difficult because of limited resources. In this paper we present a lightweight end–to–end deep learning approach for image enhancement. To improve the performance, we present mux layer and demux layers, which could perform up–sampling and down–sampling by shuffling the pixels without losing any information of feature maps. For further higher performance, denseblocks are used in the models. To ensure the consistency of the output and input, we use weighted L1 loss to increase PSNR. To improve image quality, we use adversarial loss, contextual loss and perceptual loss as parts of the objective functions during training. And NIQE is used for validation to get the best parameters for perceptual quality. Experiments show that, compared to the state–of–the–art, our method could improve both the quantitative and qualitative assessments, as well as the performance. With this system, we get the third place in PIRM Enhancement– On–Smartphones Challenge 2018(PIRM–EoS Challenge 2018).