Dehazing for Real-World Scenarios through Adaptive Perturbation Injection and Local-Global Dual-Stage Defense

Recovering high-quality clear images from hazy images is a fundamental task in the field of image processing. Although existing dehazing models have achieved success on synthetic hazy images, they still face significant challenges when dealing with real-world hazy images. To address this, we propose an adaptive perturbation injection and dual-layer defense method for real-scene image dehazing, composed of a perturbation generation module, a local-global dual-layer defense module, and a semantic guidance module. The perturbation generation module adaptively generates perturbation cues to simulate the degradation factors during the formation of real-scene hazy images, and injects these perturbations into synthetic-domain hazy images to reduce the distribution gap between synthetic and real-domain hazy images. The local-global defense module resists perturbation information from both local and global perspectives, enhancing the dehazing performance of the model in real scenarios. To address the issue that the dual-layer defense mechanism may damage useful image features during dehazing, a semantic guidance module is introduced, which incorporates semantic information into image reconstruction to further improve the quality of dehazing results. Experiments on the public RTTS dataset demonstrate that the proposed method achieves visually better dehazing results in real scenes, improves FADE by 0.137 to 1.361, and increases BRISQUE by 4.929 to 20.944 compared to mainstream methods. The code for this method has been open-sourced and is available at https://github.com/songshuaitian/API-Net.