Lightweight progressive recurrent network for video de-hazing in adverse weather conditions

Automated outdoor vision-based applications have witnessed a growing demand in everyday life. These applications, ranging from surveillance and traffic management to environmental monitoring and autonomous systems, utilize computer vision and imaging technologies to automate various processes. However, adverse weather conditions such as fog, snow, rain, and haze can severely impair video quality, limiting the performance of automated applications. Therefore, the pre-processing techniques like video de-hazing can improve visibility, ensuring that downstream applications have better performance and reliability. Further, the existing methods rely on computationally heavy networks for the task of video de-hazing. In this work, a lightweight ( \(\sim \) $\sim$ 0.24M parameters) progressive recurrent network for video de-hazing is proposed. In that, a multi-kernel feature block is proposed that addresses the challenge of varying haze intensities within frames, allowing the network to adapt and selectively emphasize features based on local haze characteristics. Additionally, a multi-receptive edge extraction block is proposed that enhances edge information by leveraging features learned with different receptive fields, capturing both local and global information. Experimental analysis on various video de-hazing datasets proves the significance of the proposed method over state-of-the-art methods for video de-hazing task. Testing code is provided in the supplementary material.