Enhancing visibility in hazy conditions: A multimodal multispectral image dehazing approach

Abstract

Improving visibility in hazy conditions is crucial for many image processing applications. Traditional single-image dehazing methods rely heavily on recoverable details from RGB images, limiting their effectiveness in dense haze. To overcome this, we propose a novel multimodal multispectral approach combining hazy RGB and Near-Infrared (NIR) images. First, an initial haze reduction enhances the saturation of the RGB image. Then, feature mapping networks process both the NIR and dehazed RGB images. The resulting feature maps are fused using a cross-modal fusion strategy and processed through convolutional layers to reconstruct a haze-free image. Finally, fusing the integrated dehazed image with the NIR image mitigates over/under exposedness and improves overall quality. Our method outperforms state-of-the-art techniques on the EPFL dataset, achieving notable improvements across four key metrics. Specifically, it demonstrates a significant enhancement of 0.1932 in the FADE metric, highlighting its superior performance in terms of haze reduction and image quality. The code and implementation details are available at https://github.com/PaulMahdizadeh123/MultimodalDehazing.