Real-Time Video Dehazing for Industrial Image Processing

In today’s industries, automation, reliability, robustness and accuracy are pivotal problem to cut costs and increase productivity and quality. Visual sensor networks are vital control and monitoring tools for continues, on-line imaging and real time image processing in production and plant process. Most of the industrial videos are captured in hazy weather and usually degraded by suspended particles of atmosphere, such as smoke, fog, rain, and snow, which limits the visual quality of image. This hinders the ability of artificial intelligent driven systems to achieve automation, reliability and accuracy. Recovery of the clear visuals from the input hazy videos is challenging problem. Instead of relying on explicitly estimating the key component of atmospheric scattering model, we present end-to-end CNN model, which directly recovers the clear images from hazy images. This end-to-end architecture makes it an ideal pre-processing tool into other deep models for increasing the efficiency of various computer vision tasks in real time systems, such as Retina-Net for object detection, ResNet for object recognition. Experimental results demonstrate the effectiveness and robustness of proposed framework by outperforming the stat-of-the-art approaches in terms of time complexity and visual quality.