DDNet: Detection-Focused Dehazing Network

Currently, deep learning-based object detection has achieved significant success on traditional datasets. However, adverse weather conditions such as heavy hazy can greatly degrade the performance of object detection. To mitigate this problem, we propose a novel Detection-focused Dehazing Network (DDNet). This architecture utilizes low-level image restoration to assist in high-level visual tasks, enabling end-to-end training and inference. Specifically, we design an Rep-Inception model based on structural re-parameterization to enhance feature extraction speed. To address the problem of feature loss in hazy scene images, we design a dual-branch fusion network with dehazing module and attention fusion module, effectively integrating original features under hazy conditions with features generated after dehazing. After conducting numerous experiments on the RTTS and VOC Foggy datasets, the outcomes are promising and show the efficacy of our method.