PTDNet: Parallel text-guided distillation network for low-light image fusion

Abstract

A common issue in low-light image fusion is overexposure during the daytime and overexposure of point light sources at night. Existing brightness enhancement networks used in low-light image fusion often lost significant information, due to the lack of supervision from high-brightness images. As a result, it is difficult for the existing low-light image fusion networks to handle complex, low-light environments. To address the challenge of enhancing image details while avoiding information loss, a novel image enhancement and fusion network (PTDNet) is proposed in this paper. PTDNet combines text guidance and knowledge distillation techniques to enhance image details and preserve information in low-light conditions. Moreover, PTDNet employs parallel CNNs and Mamba-based feature extraction and fusion modules, in order to effectively handle overexposure issues in low-light environments while preserving the accuracy and naturalness of image details under various lighting conditions. Therefore, PTDNet cannot only overcomes the brightness inconsistency issues found in traditional methods, but also enhances image visibility and clarity in low-light conditions. In the experimental section, PTDNet was qualitative and quantitatively validated on three datasets. The qualitative experimental results show that PTDNet effectively addresses the overexposure problem in traditional methods, significantly improving image quality and making details more transparent and naturally visualized. The quantitative experimental results indicate that PTDNet performed better on key metrics such as AG, EN, SF, and SD for the LLVIP and MSRS datasets.