MPE-DETR: A multiscale pyramid enhancement network for object detection in low-light images

• Published in: Image and vision computing 2024-10, Vol.150, p.105202, Article 105202
• Main Authors: Xue, Rui, Duan, Jialu, Du, Zhengwei
• Format: Article
• Language: English
• Subjects: Computer vision; Low-light images; Multiscale pyramid networks; Object detection
• ISSN: 0262-8856
• DOI: 10.1016/j.imavis.2024.105202

Object detection has broad applications in areas such as autonomous driving, security surveillance, and deep-sea exploration. However, the performance of current detection algorithms significantly decreases due to the loss of detail, increased noise, and color distortion in images under low-light or nighttime conditions. To address this problem, we propose a plug-and-play multiscale pyramid enhancement network (MPENet), which elegantly cascades with RT-DETR to establish an end-to-end framework for low-light object detection, named MPE-DETR. First, MPENet utilizes Gaussian blur to decompose images into Gaussian pyramids and Laplacian pyramids at different resolutions. Specifically, we designed a high-frequency texture enhancement (HTE) module to capture the edge and texture information of images, and a low-frequency noise smoothing (LNS) module to better understand the overall structure of images and capture global-scale features. Additionally, by concatenating the output features of the HTE and LNS modules along the channel dimension, feature fusion across different scales is realized. We conducted experiments on the ExDark and ExDark + LOD datasets, which are designed for low-light object detection. The results indicate that the proposed method achieved an improvement of 2.1% in mAP@.5 compared to that of existing SOTA models on the ExDark dataset, and demonstrated strong generalizability and robustness on the ExDark + LOD dataset. The code and results are available at https://github.com/PZDJL/MPENet. •A plug-and-play multi-scale pyramid enhancement network (MPENet), for low-light object detection.•The method proposed in this paper outperforms existing methods in terms of detection accuracy and robustness.•The proposed model performs well under low light conditions and is suitable for real-time applications.