Dynamic Low-Light Image Enhancement for Object Detection via End-to-End Training

摘要

Object detection based on convolutional neural networks is a key area in computer vision. The illumination component in images significantly affects detection performance, especially under low-light conditions. Although low-light image enhancement can improve image quality and detection performance, existing methods may negatively affect some samples, making it difficult to improve overall detection accuracy in low-light environments. This paper proposes a novel framework that combines low-light enhancement with object detection, enabling end-to-end training. The framework dynamically selects the appropriate enhancement subnetworks for each sample to improve detector performance. The approach consists of two stages: the enhancement stage, which enhances low-light images based on various enhancement methods and outputs corresponding weights, and the detection stage, where the weights provide information for object classification to generate high-quality region proposals, resulting in more accurate detection. Experimental results demonstrate that the proposed method significantly improves detection performance in low-light environments.

巫义锐

青年教授, CCF 高级会员

My research interests include Computer Vision, Artifical Intelligence, Multimedia Computing and Intelligent Water Conservancy.