Plug-and-Play Robust Aerial Object Detection Under Hazy Conditions

Abstract

Object detection under hazy weather conditions is a huge challenge due to the serious obscuration by fog. Recently, to improve detection accuracy in inclement weather, some excellent specifically designed algorithms for foggy conditions have been presented to identify objects from hazy images. However, applied to the detection task in hazy weather, state-of-the-art (SOTA) solutions have to change the architectures and training parameters of existing widely used original object detectors for clear weather, resulting in the waste of resources. To deal with this problem, in this work we propose a plug-and-play aerial object detector under foggy conditions, without the need of altering the structures and retraining the parameters of original task networks at all. The proposed algorithm includes an image enhancement and fusion branch working with a well-known SOTA object detection network for fine weather conditions, where this original detector is completely intact. The novel branch consists of a haze feature elimination module and an object feature restoration module, which are developed to remove fog information and recover clean object feature pyramid, respectively. Moreover, we also propose a detection feature pyramid loss function to more accurately restore object features. Experimental results show that the proposed plug-and-play aerial object detection framework performs well in not only fine weather but hazy weather conditions, while remaining the architecture and parameters of the original detection task network unchanged. This proposed flexible framework can be adapted to adverse weather conditions in all situations.

Keywords

• Aerial object detection
• clear weather
• hazy weather
• training parameters