Half Aggregation Transformer for Exposure Correction

Ziwen Li, Jinpu Zhang, Yuehuan Wang

Abstract: Photos taken under poor illumination conditions often suffer from unsatisfactory visual effects. Recently, Transformer, avoiding the shortcomings of CNN models, has shown impressive performance on various computer vision tasks. However, directly leveraging Transformer for exposure correction is challenging. On the one hand, the global Transformer has an excessive computational burden and fails to preserve local feature details. On the other hand, the local Transformer fails to extract spatially varying light distributions. Both global illumination recovery and local detail enhancement are indispensable for exposure correction. In this paper, we propose a novel Half Aggregation Transformer (HAT) architecture for exposure correction with a key design of Half Aggregation Multi-head Self-Attention (HA-MSA). Specifically, our HA-MSA establishes inter- and intra-window token interactions via window aggregation and window splitting strategies to jointly capture both global dependencies and local contexts in a complementary manner. In addition, we customize an illumination guidance mechanism to explore illumination cues and boost the robustness of the network to handle complex illumination. Extensive experiments demonstrate that our method outperforms the state-of-the-art exposure correction methods qualitatively and quantitatively with cheaper computational costs.