Dual-stream feature aggregation and dual guided upsampling for efficient multi-exposure correction

Improperly exposed images significantly degrade visual quality, limiting their applicability in machine learning tasks and real-world scenarios. To address this issue, multi-exposure correction methods have been developed to enhance images under diverse exposure conditions. However, existing approaches involve complex architectures, resulting in high computational costs and a large number of parameters, making them impractical for real-time applications. To overcome these limitations, we propose an efficient multi-exposure correction network, called EMECNet. EMECNet efficiently extracts a luminance feature for exposure correction and a detail feature for preserving fine details from low-resolution images, and effectively aggregates them with the guidance of exposure information. Subsequently, the aggregated features are upsampled by the proposed dual guided upsampling, which explores intensity and sub-pixel information to achieve accurate high-resolution restoration. Experimental results on the ME and SICE datasets show that EMECNet achieves the lowest computational costs and parameter counts, outperforming the state-of-the-arts in both efficiency and performance.