Low-Light Detector Based on Feature Filtering and Enhancement

Abstract

Low-light scenarios pose a challenge for object detection due to noise and low brightness. The degraded feature maps are the main factors affecting detection performance. To improve the quality of low-light feature maps, we propose double adaptive filters (DAF) and dark feature pyramid network (DFPN). DAF enhances feature representation by filtering and enhancement. Applying the adaptive kernel adjustment to the filters effectively improves interactions in the current dimension. Due to the limited detail information in low-light feature maps, the DFPN is proposed to compensate for the details. We integrate these two modules into the advanced YOLOX detector. We achieve state-of-the-art performance with unloaded and loaded COCO weights, achieving 69.1% and 83.2% accuracy on the ExDark dataset.