Enhancing low-light instance segmentation through feature-level denoising

About

Using our proposed weighted non-local blocks (wNLB) we show that their feature denoising enhances performances in instance segmentation and tracking in low-light conditions.

Funders

MyWorld Strength in Places

Links

Abstract

Instance segmentation accurately delineates the precise boundaries of each distinct object in an image or video. However, performing this task in low-light conditions presents challenges due to issues such as shot noise from low photon counts, color distortions, and reduced contrast. In this work, we propose a plug-and-play solution designed to address these complexities. Our approach integrates weighted non-local blocks (wNLB) into the feature extractor, enabling inherent denoising at the feature level. The proposed method incorporates learnable weights at each layer, allowing the network to adapt to the varying noise characteristics across different feature scales. We demonstrate that our wNLB improves the performance of object detectors and trackers when compared to pretrained networks.

Citation

If you use our work in your research, please cite using the following BibTeX entry:

@inproceedings{10.1117/12.3054001,
  author = {Joanne Lin and David R. Bull and Nantheera Anantrasirichai},
  title = {{Enhancing low-light instance segmentation through feature-level denoising}},
  volume = {13460},
  booktitle = {Machine Learning from Challenging Data 2025},
  editor = {Panagiotis  (Panos) Markopoulos and Bing Ouyang and George Sklivanitis},
  organization = {International Society for Optics and Photonics},
  publisher = {SPIE},
  pages = {134600A},
  keywords = {Segmentation, Recognition, low light, non-local blocks , denoising},
  year = {2025},
  doi = {10.1117/12.3054001},
  URL = {https://doi.org/10.1117/12.3054001}
}