Algorithm Improvement, Parallel Implementation And Optimization Of Dehazing Using Dark Channel Prior

The method of image dehazing is a significant issue concerned by both digitalimage process and computer vision. Dehazing method can transparently improvesdefinition of scenery in an foggy image, and correct the color distortion caused by thefog in the air. Dehazing technology is the basis of many computer vision algorithms,because most computer vision algorithms can only deal with clear image, and the fogshould be removed before applying these algorithms to foggy image. Dehazingalgorithms are widely required in common life, for example, they can help people toachieve clear pictures and information from photos taken in foggy weather. However,the current dehazeing algorithms aren’t able to satisfy the application demands, andneed to research and improve more deeply.The method of image dehazing using dark channel prior is one of the dehazeingmethods with good effects and owns a number of advantages, but its severaldisadvantages limit the extent of its applications, which include bad effect to objectivesimilar to atmosphere light, heavily memory-consuming and high time complexity. Inthis paper, we base on the algorithm of image dehazing using dark channel prior, andimprove the algorithm to overcome its defects. Our algorithm improvements include:(1)propose new method through recognizing objective similar to atmosphere light toimprove dehazing effect;(2) introduce the method of “guide filter” to eliminate thedisadvantage of heavily memory-consuming;(3) bring forward novel way named“multilevel blocks” to reduce computational complexity in module of computing darkpixel and computing initial transmission;(4) put forward integrated evaluation value todecrease computational complexity in module of selecting atmosphere light. Thesealgorithm improvements make the dehazing algorithm using dark channel prior obtainbetter effect and be applied to wider application extent.In recent years, single CPU core has arrived its bottleneck of performance andefficiency, The research on multi-core, cluster and GPU become the trend of improvingoverall performance and accelerating the application, so parallelism becomes animportant method of algorithm research and acceleration. By now, the parallelismresearch of dehazing algorithms is few, but the effect of reducing execution time only byalgorithm improvement is limited, so we must resort to parallelism to largely diminishthe execution time. During the process of parallel implementation, we excavate the data parallelism and partition the task properly, so as to achieve basic GPU parallel programwhich obtains about7x acceleration. After that, depending on the characteristics of thealgorithm and GPU platform, we propose some new parallel method, such ascalculating accumulative sum with keeping intermediate results, to develop theparallelism. We also utilize several optimization technologies, such as combination ofkernel and skillful use of shared memory, to reduce the overheads of correspondenceand startup during algorithm execution, and obtain optimized parallel algorithm, andachieve20x-30x acceleration. Finally, we can process the image with resolution600×400at the speed of25frames per second by our optimized parallel algorithm.