Lightweight Image Enhancement Network for Mobile Devices Using Self-Feature Extraction and Dense Modulation

Recently, as the use of media services including Social Networking Service (SNS) has been increased, customers who transmit and share their images and videos through the online platforms by using mobile phones also have increased. Since these images are processed with downsampling and compression through the online platform server, their visual quality is noticeably degraded when stored from the server. In addition, the cases tend to increase that users capture their oldphotographs by using mobile phone cameras and store them. It causes the quality worsening, since most old-photographs have some degradation characteristics such as faded color and irregular noise patterns, etc. Therefore, there are increasing needs for image enhancement solutions, which reconstruct a high-quality image from a low-quality image. Such solutions are operated on the mobile device requirements, and can be applied for image quality enhancement including face renovation and old-photograph restoration. As depicted in Fig. 1, image enhancement solutions perform not only super-resolution with compression artifacts removal for low-resolution images, but detail improvement including color enhancement while maintaining the resolution. Many studies have introduced deep convolutional network based image super-resolution methods [1-3]. It is obvious that CNN has become a disruptive technology and noticeably changed the landscape of image processing applications [4], [5]. They exhibit significant improvements for semantic fidelity and visual reality of reconstructed images. However, these solutions have limitations to perform in lightweight models for SoC requirements while keeping identities of their networks. In spite of increasing demands of image enhancement methods which enable to operate in ondevice environments, conventional lightweight networks from existing methods show low performance with undesired artifacts, such as aliasing effects. Lightweight Image Enhancement Network for Mobile Devices Using Self-Feature Extraction and Dense Modulation