Learning to Restore Fire Smoke Monitoring Images: A Dataset and A Transformer-Based Method

Image desmoking is an important task in computer vision, aimed at recovering clear visual content from images affected by smoke. Smoke causes reduced contrast and loss of details, which negatively impacts subsequent image analysis and understanding tasks. Traditional desmoking methods often rely on physical models or prior knowledge, but they are limited when handling complex scenes and non-uniform smoke. With the development of deep learning, CNN-based methods have made significant progress. However, these methods focus on local features, potentially neglecting the global contextual information. Recently, Transformer models have shown unique advantages in desmoking tasks due to their self-attention mechanism. Transformer-based models can capture both local features and global context, improving desmoking performance. This paper proposes a Transformer-based desmoking method called DesmokeFormer. This method focuses on both the details and global context of the image during the desmoking process, thereby enhancing the desmoking effect. To validate the effectiveness of DesmokeFormer, we construct the Desmoke6k dataset, which includes hazy and clear image pairs from various indoor fire smoke scenarios. Experimental results demonstrate that DesmokeFormer outperforms existing state-of-the-art desmoking methods in terms of image quality and perceptual evaluation, proving its potential for practical applications.