A lightweight adaptive feature-domain defogging dynamic convolutional network for object detection in foggy conditions

Rapid and accurate Advanced Driver Assistance Systems (ADAS) Object Detection in foggy environments is crucial for autonomous driving tasks. Existing methods typically enhance images before detection, which compromises real-time performance. Popular detectors also rely on stacks of static convolutions for feature extraction and fusion, making it hard to capture details in fog-degraded images and reducing accuracy. We therefore propose a lightweight Adaptive Feature-domain Defogging Dynamic Convolutional Network (AFD-YOLO) for object detection in foggy scenes. AFD-YOLO integrates a flexible convolutional architecture and dynamic scale selection throughout training, while adopting lightweight designs across the backbone, the Feature Pyramid Network, and the detection head. First, a Context-Guided Downsampling (CGD) module suppresses haze-induced scattering and restores discriminative details in the feature domain, thereby replacing complexity-increasing preprocessing steps. Second, to alleviate texture diffusion, a Cross Stage Partial Parallel Groupwise Multi-scale Convolution (CSP-ParGMC) module uses parallel multi-scale hybrid convolutions to adaptively aggregate cross-scale information from shallow layers. To address attenuation of structural orientation, AFD-YOLO introduces a Cross Stage Partial Dynamic Kernel Mixture Bottleneck that employs hierarchical kernel configurations and Dynamic Kernel Convolution (DKConv2d) for Scale and Orientation Selection; together with an Efficient Bidirectional Lightweight Feature Pyramid Network (EBLFPN), the flexible architecture transmits diverse gradients to the detection head with lower computation. Finally, experiments on synthetic and real-world foggy datasets demonstrate the effectiveness and efficiency of AFD-YOLO for real-time deployment. On the RTTS dataset, AFD-YOLO reduces parameters by 36% while improving mAP@0.5 by 3.08%.