Research On Fog Image Restoration Based On Convolutional Neural Network

Recently,with the continuous development of big data and artificial intelligence,many intelligent systems based on information processing have gradually increased the requirements for input image quality and clarity.For example,the traffic control system and the driverless system.Due to the deterioration of air quality and the increasing frequency of haze weather,the actual image quality is generally low.Therefore,improving the quality of foggy images has urgent practical needs and broad application prospects.In this paper,the fog image restoration technology and its quality evaluation are discussed in depth,and the following research is carried out:Firstly,a method of quantitatively estimating image fog using dark channel is proposed.Then an estimation algorithm combining color analysis and brightness analysis is proposed.Compared with the previous methods which only focus on brightness estimation,this method combines the estimated color similarity in YCbCr space to select representative fog pixel candidate points for atmospheric light value calculation.The objective of accurate and optimal estimation of atmospheric light value is achieved.Secondly,the defogging convolutional neural model is designed according to the dark channel priority characteristics of the image.The model was developed based on dark channel characteristics and was derived from the decontamination algorithm.The projected map of the fog image can be accurately predicted via the optimized network model,and Restoring defogging image according to physical atmospheric scattering template.Finally,the quality of defogging images is evaluated from both subjective and objective aspects.Faster-RCNN target recognition is used to compare the number of targets identified by different methods before and after defogging.The test results show that the proposed algorithm not only has good fog removal effect,but also contributes to image target recognition and classification and other subsequent image processing work.