DEO-NET: A Small Object Detection Method for Foggy Low-Visibility Scenarios

Abstract

In foggy and other low-visibility conditions, atmospheric scattering and light attenuation severely degrade image contrast and edge clarity, leading to weakened feature discriminability and degraded object detection performance, especially for small targets in complex backgrounds.To address these challenges, this paper proposes DEO-NET, a fog-aware object detection framework tailored for transmission line insulator defect detection. Built upon the YOLO11 architecture, DEO-NET introduces coordinated enhancements in feature extraction, feature fusion, and channel modeling. Specifically, a Detail-Enhanced Convolution (DEConv) is employed to strengthen local gradient variations and edge responses under low-contrast conditions. Meanwhile, an Online Reparameterized Feature Fusion (OREPA) module is integrated to optimize multi-scale feature aggregation and improve feature representation completeness. Furthermore, an Efficient Squeeze-and-Excitation (ESE) mechanism is adopted for adaptive channel reweighting, effectively suppressing fog-induced background interference and reinforcing the network’s capability to capture critical defect features. To validate the effectiveness of the proposed method, comprehensive comparative experiments and ablation studies are conducted on both a synthetically fog-degraded dataset of transmission line insulator defects and a real-world foggy natural-scene dataset. Experimental results demonstrate that, on the synthetic foggy insulator defect dataset, DEO-NET achieves an mAP@0.5 of 0.669representing a 28.4% relative improvement over the baseline YOLO11-n (0.521). On the real-world foggy dataset, its mAP@0.5 increases from 0.721 to 0.766 (a 6.2% absolute gain), accompanied by concurrent improvements in Precision and Recall. Notably, DEO-NET attains superior overall performance and enhanced robustness across diverse foggy scenarios, while maintaining acceptable model complexity in terms of both parameter count and computational cost.