Presentation description

ClaraVisio advances image defogging through Image-to-Image translation (I2I), a machine learning technique that transforms foggy images into clear ones. This technology is particularly useful for applications such as autonomous vehicles, which face challenges in foggy conditions. Traditional I2I models are limited by scarce and difficult-to-acquire fog datasets.

Our approach builds on existing models like StereoFog and FogEye by introducing a novel dataset of free-floating fog. We constructed a custom data collection setup using a Raspberry Pi and a 3D-printed cone to ensure consistent fog conditions during image capture. The Pix2Pix model was trained on over 1,000 pairs of foggy and clear images, collected under diverse conditions including varying fog densities and weather scenarios.

This method contrasts with previous research that used controlled fog environments by incorporating natural fog variations. The model achieved notable performance metrics, including an average structural similarity index measure (SSIM) of over 75%, which evaluates the quality of the reconstructed image against the actual clear image. The final real-time defogging model demonstrates the effectiveness of this approach for improving visibility in fogged images, with potential applications in autonomous driving systems.

ClaraVisio enhances defogging technology and contributes to the field of image-to-image translation by addressing previous limitations and enabling practical, real-time applications. Despite the relatively small dataset, this work lays the groundwork for further data collection and model refinement.