Residual Feature Distillation Channel Spatial Attention Network for ISP on Smartphone

Abstract

With the increasing popularity of mobile photography, more and more attention is being paid to image signal processing(ISP) algorithms used to improve various perceptual aspects of mobile photos. For this, a learned smartphone ISP task was proposed to develop an end-to-end deep learning-based ISP pipeline that can replace classical hand-crafted ISP and imitate target RGB images captured by digital single-lens reflex camera(DSLR). However, hardware limitations of mobile phone make it essential and challenging to achieve an acceptable trade off between computational cost and performance. In this paper, a light-weighted and powerful network named residual feature distillation channel spatial attention network (RFDCSANet) is proposed for end-to-end learned smartphone ISP task. To be specific, we employ modified residual feature distillation block(RFDB) including channel spatial attention(CSA) mechanism to progressively refine distilled features and adaptively fuse channel and spatial features. Particularly, we utilize a re-parameterizable block, namely edge-oriented convolution block(ECB) as the basic module to improve performance without introducing any additional cost in the inference stage. The proposed solution ranked $\boldsymbol{3^{rd}}$ 3 rd in Mobile AI 2022 Learned Smartphone ISP Challenge (Track 1) with $\boldsymbol{1^{st}}$ 1 st place PSNR score.