The Research And Actualization On The Optimization Of The Haze Removal Algorithm Based On Convolutional Neural Network

The haze removal algorithm has been wide used under the smoggy circumstances.Now the haze removal algorithm based on the dark channel prior achieves a good performance.But the long processing time,existence of texture in the sky,unnatural cohesion,dark results limit its further application.With the development of deep learning,it`s a trend to combine the deep learning algorithm with the haze removal demand like using CNN(Convolutional Neural Network)to deal with the smoggy images.In this paper,the haze removal algorithm based on dark channel prior is optimized and a haze removal algorithm based on convolutional neural network is proposed.Meantime,OpenCL(Open Computing Language)is used to design and optimize the parallel convolutional neural network.Training CNN with parallel method takes much less time and higher order invariant features from the original input characteristics can be learned.Meantime,the following problems can be solved including the low hard disk utilization rate,the high cost,the difficulty of hardware actualization and the cross-platform issues.Besides,the requirement that shorten the time of haze removal processing is met.First of all,the concepts,principles and key technologies of convolutional neural network are introduced.Taking typical LeNet5 network as an example,the paper describes the functions and actualization of each layer.Then the haze removal algorithms,mainly the dark channel prior,are presented.Based on that,four points to optimize it are proposed.(1)Compute atmospheric light value based on the "white zone".Even a small heavy fog area is contained in the input image,a clear result can be achieved.The improvement can reduce the computing time,too.(2)Use a different image formula to recover the picture.we only need to compute atmospheric light value and do not need the transmittance.That takes much less time.(3)An algorithm based on sky segmentation is used to reduce the textures of sky.(4)Color correction based on Just Noticeable Difference(JND)theory is used to brighten the result.Second,an algorithm based on the dark channel prior algorithm to add haze to the normal pictures is proposed.Then label the haze images with haze concentration.The data set is used to train and test the haze removal CNN.Third,a haze removal algorithm based on the convolutional neural network is realized.Firstly,the algorithm classifies the haze image according to haze concentration,then finds out the sky part of the picture and deals with Contrast Limited Adaptive Histogram Equalization(CLAHE),the rest part is handled by improved dark channel prior algorithm with the corresponding parameters of the non sky part.Finally,brighten the whole picture by JND.The result is better than the one treated by dark channel prior algorithm.Fourth,a serial and a parallel CNN using OpenCL based on VS2015 are designed and realized.Two parallel optimization schemes to speed up the training process onetwork are made.After the test,the convolutional calculation part takes the longest.Therefore,a convolution kernel based on parallel batch data is designed to accelerate network training and Shared Virtual Memory(SVM)is used to decrease the exchanging time between the host and the devices.In the end,the improved parallel CNN takes 3.4% training time corresponding the one trained in serial mode.In my work,the dark channel prior algorithm optimization scheme is put forward,a designed convolutional neural network for haze elimination is accomplished,and the training process of the net is accelerated with OpenCL All the work has significance for the haze removal algorithm and the parallel acceleration of the convolutional neural network.