GLINet: Global-Local Interacting Network for Object Detection Under Inclement Weather Conditions

Recently, deep learning-based object detection methods have achieved promising success in normal scenarios. However, the performance of such methods often fails to locate objects from the degraded images captured under inclement weather conditions. The existing methods mostly utilize image enhancement techniques or domain adaptation methods, which focus on local improvement and ignore global information. To tackle this issue, we propose a novel Global-Local interacting network (GLINet) framework to employ the interplay between global features and local features. Specifically, a Global Enhancement Module (GEM) and a Local Enhancement Module (LEM) are designed to enhance the degraded images parallelly. For the GEM, we design a Fourier Branch and a Spatial Branch, which performs global enhancement in the Fourier domain and complements the Fourier Branch, respectively. For the LEM, we design a Vision Encoder to forecast a parameter which is employed on a sharpen filter to emphasize the local detail and texture information. we combine GLINet with YOLOv3 in an end-to-end manner. The experimental results demonstrate significant improvement in detection performance over several methods in both foggy and low-light scenarios.