Research On The Image Dehazing Algorithm Based On OpenCL

Image dehazing represents a hot research area in image processing in recent years,which boasts vital application value in ITS(Intelligent Transportation System), military,monitoring system and remote sensing system,etc. Currently, there are two computingmethods for image dehazing: one is based on image enhancement; the other is based onimage restoration. Image enhancement does not need to establish physical model andthe image detail regions are highlighted by increasing contrast, the algorithm of whichis relatively simple. However, its dehazing effect is not as good as image restoration.Most of algorithms based on image restoration boast excellent dehazing effect. But it ishighly complicated and carries massive calculation and takes a long time, dark channelhypothesis is a good example for this.Recently, the dark channel hypothesis algorithm research focuses mostlyconcentrated in the dehazing algorithm and the dehazing effect to image, the algorithmaccelerated research is still very limited, and this algorithm has a very longtime-consuming process, by which is difficult to be widely applied in engineering. Forthis situation, this paper focuses on algorithm acceleration, through the use of parallelprogramming standards and OpenCL GPU parallel architecture to realize the darkchannel hypothesis algorithm parallel speedup，which would be to lay the foundationfor project realization in the future.Firstly, this thesis studies parallel computing technique. The parallel computingprocess and merits and demerits of parallel architecture are concluded by analyzingand comparing current parallel computing methods and equipments, then analyzes andsummarizes the GPU architecture, which lays a solid foundation for the followingparallel computing methods.Secondly, this thesis thoroughly analyzes dehazing computing method based ondark channel hypothesis. This method is based on atmospheric physics model,including attenuationterm and atmospheric optic term. Based upon the abovementioned theories and research, blurred image model in foggy days is concluded andthe solving process of dehazing computing method based on dark channel hypothesis isdemonstrated in view of atmospheric optic A and transmissivity t(x) in blurred model.In addition, this thesis carries out comparison and analysis for the methods of refiningtransmissivity t(x) thus summarizes methods suitable for parallel realization andsystem function of solving image dehazing.Finally, this thesis studies the programming norms and framework of OpenCL, based on which, the degree of parallelism in every step of dehazing computing methodbased on dark channel hypothesis is carefully analyzed, and improves and optimizesthose steps unsuited for parallel disassembling. It also makes use of OpenCL C toperform parallel realization and speed up computing process by optimizing register inGPU.