Towards Deep Residual Learning Atmospheric Veil For Single Image Dehazing

Based on current trends in image processing and artificial intelligence,it has become an increasingly growing trend for research to be aimed at minimizing computational cost in terms of memory management and essentially computational time.This is as a result of real time requirements for image processing in several computer vision applications ranging from medical imaging and computer assisted interventions to instance and semantic segmentation systems that apply object detection for autonomous systems.Given the need for clear vision in several tasks,image dehazing has become salient to several vision tasks such as object detection,etc.,however,several of the existing approaches perform significantly well but do that over a relatively long span of time.This is as a result of dependence of several computations such as transmission map,airlight estimation and time wise refinement of the transmission maps before dehazing tasks can be executed.In this thesis,we introduce a transmission map independent approach to dehazing,thereby eliminating the need to estimate more than required parameters for dehazing.Furthermore,in an attempt to improve on the state of the art in terms of computational run time and quality of scene radiance recovery,we propose a deep residual neural network that speeds up training time significantly due to its mode of internal skip connection and its feedback connections respectively.Finally,we aim to improve dehazed image quality by maintaining depth information and a close replication of the original scene radiance.