Diverse multi-scale features absorption for lightweight object detection models in inclement weather conditions

Abstract

In recent years, numerous lightweight object detection models have been introduced and successfully deployed on low-computation devices. However, these models mainly focus on detecting objects in favorable weather conditions and do not adequately account for inclement conditions, particularly in the presence of fog. This significantly leads to the drastic performance degradation of object detectors, primarily attributable to the decreased visibility. To tackle the aforementioned deficiency, we introduce a novel diverse multi-scale feature absorption network (DMFA-Net) to guide lightweight detectors work efficiently in foggy weather conditions. Our approach achieves its objective through the close collaboration of three subnetworks: a detection enhancement subnetwork, a depth mining subnetwork, and a lightweight detection subnetwork. The lightweight detection subnetwork achieves a significant accuracy improvement by absorbing and learning a range of useful features from both the detection enhancement and depth mining subnetworks through diverse multi-scale feature absorption loss. Extensive experiments demonstrate that our DMFA-Net effectively boosts baseline lightweight detectors in accurately localizing and classifying objects, without adding any computational cost. Additionally, it outperforms representative competing approaches on both synthesized and real-world foggy image datasets.