Dark Channel Difference Driven Dehazed Image Quality Assessment With Progressive Injection

Image dehazing, as a typical representative of low-level vision tasks, holds significant importance, prompting the proposal of numerous dehazing methods. Given that none of these methods can perfectly restore all hazy images, developing robust objective assessment methods remains a critical research priority in this field. Most existing objective assessment methods for dehazed images fail to fully leverage haze-related features and instead rely on basic dehazed images alone. Additionally, they often neglect effective feature fusion, which severely limits model performance. To address these issues, we propose a dark channel difference driven dehazed image quality assessment with progressive injection (DCPI-DQA). First, a dark channel difference map is generated using both the dehazed image and the hazy image. Then, a two-stream multi-scale feature extraction network is employed to extract haze density features and dehazing-aware features. During the feature extraction process, haze density features are progressively injected into the dehazing aware features using the proposed progressive injection mechanism. Finally, multi-level fusion is applied for feature integration, and the quality score is obtained through quality regression. Experiments conducted on four publicly available dehazed image datasets demonstrate that our method achieves state-of-the-art performance. Furthermore, cross-dataset experiments indicate that the proposed method exhibits strong generalization. Code and resources are available at https://github.com/BoHu90/DCPI-DQA.