UDC Image Restoration via Lightweight Dual-Domain Network

Abstract

With the development of full-screen devices, smartphone manufacturers have started to explore placing the camera behind the screen, offering larger display-to-body ratios. However, this under-display camera (UDC) imaging system causes varying degrees of degradation for captured RGB images. To address this issue, we propose a lightweight dual-domain network (LD2Net) for restoring UDC images. Specifically, first, the UDC image is decomposed into reflection and illumination images. Then, an amplitude-phase mutual guided block is designed to reconstruct the reflection image in the frequency domain and restore the attenuated highfrequency components. Meanwhile, we employ a multi-scale hybrid dilated convolution block to reconstruct the degeneracy of UDC images in the spatial domain. Finally, the feature maps of the spatial and frequency domain outputs are fused to recover a clear image. In addition, we propose a global perceptual loss function based on the Transformer framework to help LD2Net reconstruct clearer textures and colors. Extensive experiments show that our model can achieve superior results with relatively few parameters. The valid URL for our code is at https://github.com/zzr-idam/LD2Net.