Low-Light Image Enhancement Using Multi-Branch Deep Fusion Enhancement Network

Abstract

Deficiencies like lack of visibility, blurry images, and noise affect almost all images captured in less-than-ideal lighting conditions. To cope with this issue, we propose the multi-branch Deep Fusion Enhancement Network (DFEN) that utilizes artificially generated multi-exposure images for the image enhancement operation. The model extracts feature from various levels from the different multi-exposure images and then combine the re-calibrated features to generate an overall enhanced low-light image. The model makes use of the Feature Extraction Module (FEM) to draw out features from different levels of the image, Feature Enhancement Module (EM) to enhance the features extracted, and Feature Fusion and Re-calibration Module (FFRM) to re-calibrate enhanced features, and finally merge them into an enhanced low-light image. The proposed model was evaluated on various datasets and showed to outperform various state-of-the-art techniques significantly. Additionally, numerical results for the proposed DFEN model show it outperforms other approaches in either qualitative or quantitative metrics.