FRN: Fusion and recalibration network for low-light image enhancement

This paper proposes a Fusion and Recalibration Network (FRN) for low-light image enhancement. Firstly, The proposed method generates multi-exposure images from a single image to enhance low-light images. The proposed Feature Extraction Module (FEM) extracts multi-level features from an image. The proposed method uses Feature Augmentation Module (FAM), a U-net-like structure, to encode the multi-level features and assist in the reconstruction. The proposed Feature Fusion and Re-calibration Module (FFRM) re-calibrates and merges the features to provide an enhanced output image. The advantage of dynamically selecting features from extremely bright regions of the artificially darkened images and darker regions of the artificially brightened image results in a balanced output image. The proposed model was evaluated on various datasets and significantly outperformed most state-of-the-art techniques. Additionally, the experimental assessment shows that the proposed FRN model outperforms other quantitative and qualitative assessment approaches.