MECFormer: Multiple Exposure Correction Transformer Based on Autoencoder

Images captured with wrong exposure conditions inevitably produce unsatisfactory visual effects. Thus, multiple exposure correction has drawn much attention, which should correct for degraded images due to various wrong exposure conditions. However, the problem of handling the different nature of underexposed and overexposed images makes this task challenging. In this work, we introduce the novel multiple exposure correction transformer, named MECFormer, to tackle this problem. MECFormer consists of autoencoder, encoder, and dual-path aggregation decoder. First, the autoencoder extracts multi-scale exposure features representing the level of input exposure. Second, the encoder embeds input images into multi-scale image features. Third, the dual-path aggregation decoder sequentially restores exposures by effectively aggregating multi-scale exposure features and image features. MECFormer achieves the state-of-the art performance on two multi-exposure correction datasets. Also, we provide extensive ablation studies to show the effectiveness of the proposed components.