BiEnhancer: Bi-Level Feature Enhancement in the Dark

Zhou Teng

Keywords: Low-light image enhancement; low-light object detection; night-time semantic segmentation

TL;DR: BiEnhancer, a versatile plug-in module, enables end-to-end joint training with various ligh-level vision tasks to boost performance under low-light conditions.

Abstract: The remarkable achievements of high-level vision tasks (e.g., object detection, semantic segmentation) under favorable lighting conditions highlight the persistent challenges faced in low-light vision. Previous studies have mainly focused on enhancing low-light images to create visual-friendly representations, often neglecting the differences between machine vision and human vision. This oversight has led to limited performance improvements for high-level tasks. Furthermore, many approaches rely on synthetic paired datasets for training, which can result in limited generalization to real-world images with diverse illumination levels. To address these issues, we propose a new module called BiEnhancer, which is designed to enhance the representation of low-light images by optimizing the loss function of high-level tasks to improve performance. BiEnhancer decomposes low-light images into low-level and high-level components and performs feature enhancement. Then, it adopts an attentional feature fusion strategy and a pixel-wise iterative estimation strategy to effectively enhance and restore the details and semantic information of low-light images and improve the machine-readable representation ability of low-light images. As a versatile plug-in module, BiEnhancer supports end-to-end joint training with diverse high-level tasks. Extensive experimental results demonstrate that the BiEnhancer framework outperforms state-of-the-art methods in both speed and accuracy.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4543