RawEnhancer: Attentive Multi-Exposure Selection for Bracketing Image Reconstruction

Long Sun, Jinshan Pan, Rushi Lan, Yazhou Yao, Jinhui Tang

Keywords: Bracketing image restoration and enhancement; Raw image processing; Multi-exposure HDR imaging

Abstract: Exposure bracketing is a crucial technique for capturing high-quality images in real-world scenes. The key challenge lies in effectively fusing complementary information across bracketing exposures while considering the unique characteristics of each exposure. To address this, we develop an attentive multi-exposure selection network, dubbed RawEnhancer, built on an asymmetric encoder-decoder architecture for visual-pleasing bracketing image reconstruction. Specifically, we first present an exposure-aware selective attention (EASA) that leverages gradient magnitude and color channel intensity priors from the RAW sensor data to achieve a finer representation. To enable the model to handle multiple exposures, we construct a weight-sharing encoder with the EASA layers to efficiently encode multi-scale information of each individual exposure. These multi-scale representations are then progressively fused in the decoder part through an iterative feature selection (IFS) strategy, which adaptively integrates complementary information across multiple exposures at each scale. Experimental results show that the proposed RawEnhancer performs favorably against state-of-the-art ones on benchmark datasets in terms of accuracy and model efficiency.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3593