Effects of haze and dehazing on deep learning-based vision models

Haseeb Hassan, Pranshu Mishra, Muhammad Ahmad, Ali Kashif Bashir, Bingding Huang, Bin Luo

Abstract: Most deep-learning-based vision models are trained and tested on clear images, avoiding noisy, or hazy, images. However, these models may encounter degraded images. So, it is important to recover and enhance them using a dehazing process. Dehazing usually serves as a preprocessing step for low-, medium-, and high-level vision tasks. Therefore, this article empirically studies the impact of haze and dehazing on high-level vision tasks and considers the degree to which dehazing algorithms can improve a vision model’s performance. For this purpose, we created two synthetic hazy datasets and trained several detection and classification models on both clear and hazy images. We found that haze and fog can easily affect a vision model’s performance and observed that using dehazing directly as a preprocessing step for high-level vision tasks did not substantially improve vision model’s performance but also renders performance unreliable and unpredictable. Therefore, when developing deep vision models, the research community should maintain aspects of bad weather conditions, such as haze, mist, fog, and rain, to avoid the failure of their proposed outdoor vision models.