Road-Scene Visibility Estimation Based on ResNet-18 Dual-Branch Enhanced Gated Fusion

Abstract

Road-scene visibility estimation is a critical task in intelligent transportation systems and autonomous driving perception. To address the spatially non-uniform degradation present in foggy images and the difficulty of single-scale features in simultaneously characterizing global atmospheric conditions and local object clarity, this study proposes a visibility regression network based on dual-branch feature extraction and enhanced gated fusion. The network consists of parallel global and local branches. The global branch takes the entire road image as input, while the local branch uses a region of interest (ROI) detected by YOLOv8. Two structurally identical but parameter-independent ResNet-18 networks are employed to extract global scene features and local detail features, respectively. To explicitly model the relationship between the two branches, the absolute difference and element-wise product of the two feature vectors are calculated and concatenated with the original features to form an enhanced gating input. The gating network then generates sample-adaptive fusion weights for the two branches. Experimental results demonstrate that the proposed enhanced gating fusion method achieves MAE of 215.5, RMSE of 275.2, and R2 of 0.9946, confirming the effectiveness and superiority of the proposed approach.