Fast Range Domain Filter for Efficient Dehazing of Videos using Deep Learning

Abstract

Computer Vision enables a machine to perceive the visual world and interpret intelligently through enough processing power behind a digital camera. Larry Roberts, the father of computer vision described the process of deriving 3D information of a solid object from 2D photographs in his Ph.D. thesis (MIT, 1963), Machine Perception of Three-Dimensional Solids . Both image analytics and video analytics do come under the larger umbrella of computer vision since then. For many applications of computer vision like face recognition, defect inspection, etc. analyzing a single image may be sufficient. But for identifying the motion direction of some moving object or determining activity, analyzing temporal information is also important along with spatial information. The former is image analytics and the latter one is video analytics. newlinePractical computer vision problems are generally addressed through a combination of traditional image/ video processing algorithms and machine/ deep learning models, to ensure the best accuracy, performance and scalability. Although, many recent machine/ deep learning models are being used to solve some video analytics algorithms, by themselves. Researchers have been addressing image enhancement filtering, especially for scene quality degradation through adverse environmental effects like rain and fog. newlineThe current thesis proposes (a) an algorithm of fast bilateral filtering utilizing the property of dominant sparse color in images and (b) developed dynamic dehazing algorithms addressing fog and rain using the guided bilateral filters and deep convolutional autoencoder. Based on a novel method of temporal machine learning-based scene categorization according to mutual motion between the target object and the camera, the enhancement algorithm has been adapted dynamically. Application of the developed algorithm has been shown effective in quality of enhancement, algorithm performance and subject detection accuracy in different domains of computer vision ranging from traffic video analytics to