MD-BAN: Multi-Direction Mask and Detail Enhancement Blind-Area Network for Self-Supervised Real-World Denoising

Abstract

Recently, Asymmetric PD and Blind-Spot Network (AP-BSN) has shown effectiveness for real-world image denoising. However, when the noiserelated area is large, it uses the single-center pixel mask which cannot break the noise spatial correlation, therefore the blind spot recovered from the surrounding pixels still contains noise, resulting in obviously abnormal color spots in the denoised image. In addition, AP-BSN enlarges the receptive field by stacking multiple dilated convolutional layer (DCL), but these layers may lead to block artifacts and partial pixel detail information loss due to their interpolation and overlap operations. To address the above issues, we propose a multi-direction mask convolution kernel (MDMCK) to form a blind area to further destroy largescale spatial connection noise. We also propose a detail feature enhancement (DFE) module to supplement the detail lost by MDMCK and stacking DCL. Finally, we use a robust joint loss function to train our model, generating denoised images with clean and sharp detail while alleviating the block artifacts. Extensive quantitative and qualitative evaluations of the SIDD and DND datasets show that our proposed method performs favorably.