Research On Single Image Dehazing Algorithm

As the environment deteriorates,haze weather is becoming more and more common.When acquiring images in haze weather,part of the light reflected by the object is absorbed and scattered by the suspended particles in the air on the way to the sensor,and at the same time,part of the scattered light of atmospheric light is mixed to the light the sensor finally receive,resulting in image degradation such as contrast reduction,saturation decline,and hue shift.These degradation will seriously affect the performance of computer vision systems.Therefore,image dehazing has become an important research direction in image processing.This paper studies the single image dehazing algorithm based on the dark channel prior and deep learning,which effectively improves the efficiency and effect of the dehazing algorithms.The main work is as follows:(1)A fast single image dehazing algorithm based on dark channel prior is proposed.Firstly,the depth model is established to estimate the transmittance of the hazy image.After that,the rough transmittance is obtained on basis of the dark channel prior and its local consistency characteristic.Then the rough transmittance and the transmittance estimated by depth model are fused to achieve the correction of the sky region.Finally,the image is restored using the transmittance refined by guided filter,and the tone mapping is used to the restored image.The experimental results show that the proposed algorithm has high efficiency,which can also effectively improves the visibility and contrast of the restored images.(2)An image dehazing algorithm based on convolutional neural network and dynamic ambient light is proposed.Firstly,an image library containing the paired real hazy images and transmittance images is constructed.Then,randomly sampling from the library to obtain the paired hazy patches and transmittance patches as the training set to train the transmittance estimation network.After that,using the trained network to estimate the transmittance of the hazy image.At the same time,considering the uneven illumination of images,the dynamic ambient light is used to replace the global atmospheric light,and finally the smoothed filtered transmittance and the dynamic ambient light are used to restore the image.The experimental results show that the algorithm can not only effectively restore the images,but also significantly improve the brightness and saturation of the restored images.(3)An end-to-end image dehazing algorithm based on cycle generative adversarial networks is proposed.When constructing the training set,an image library which contains paired clear images and depth images is constructed firstly.Then,the sampling and brightness adjustment operations are performed to obtain the paired synthetic hazy patches and clear patches as the training set to train our networks.Finally,the dehazing network of the trained networks is used for end-to-end dehazing.In addition,the depth weighted loss is designed to reduce the impact of the sky region in the image patches when training networks.Experimental results show that the proposed algorithm has good image dehazing effect,and the restored images have high brightness,high definition,and rich texture detail information.