Image Enhancement And Image Dehazing Based On Deep Learning

The development of society promotes people's pursuit and yearning for highquality life,including high requirements for visual experience.Nowadays,the breakthrough of industrial technology has led to the development of all kinds of imaging equipment,but due to the influence of objective conditions,there are still many defects,so that it is difficult to meet the actual needs of human eyes for images.For example,in the case of haze weather,due to the scattering of particles in the atmosphere,the image quality is seriously reduced,the object is occluded,the color is single,and the information obtained from the image is incomplete,in the night or the lack of light source environment,due to the lack of brightness,the images obtained by the camera equipment often have problems such as low definition,low contrast,loss of detail,loss of color and so on.In order to solve the problem of obtaining images in hazy days and at night,the research of image dehazing and image enhancement technology is particularly important.At present,most of the methods of image dehazing and image enhancement are based on the imaging principle of hazy image and low brightness image,and the restored clear image is obtained by backward derivation or learning.Among these methods,the deep learning method avoids the estimation of model parameters through deep network learning,but these methods lack the consideration of haze dissipation principle,color deviation,noise and so on,resulting in the distortion of the results.Based on deep learning theory and considering the defects of noise and information loss in existing physical models,this thesis proposed a new image dehazing and image brightness enhancement method.The main contribution of this thesis are as follows:1)Aiming at the problem that the traditional Retinex based image enhancement algorithm only considers the change of image brightness and ignores the noise and color deviation in the process of image restoration,this thesis proposes an end-to-end image enhancement network based on multi information flow supplement.In this method,low exposure image is used as input,and multi stream layer is input after feature extraction.In order to avoid losing details in the process of network learning,the proposed method ensures that the size of mainstream network feature map is consistent with that of network input.In order to recover more detailed semantic information and avoid the problem of insufficient information obtained by small-scale convolution kernel,the network constructs two branches by reducing the size of feature graph to provide supplementary information for the mainstream network.Finally,in order to avoid the color distortion caused by the enhancement process,the color difference loss is introduced to measure the color difference between the enhancement result and the reference image.The experimental results show that the proposed method can get good brightness enhancement results,which proves the effectiveness of the method.2)In view of the phenomenon that the haze vanishes gradually from thick to light in the real scene,this thesis proposes a progressive dehazing network to realize image dehazing.Compared with the existing dehazing methods based on end-to-end network,this thesis divides the dehazing process into two steps,which are thick fog and light fog.In the preliminary module,this thesis designs a multi-level residual mode to remove the fog step by step,so as to avoid excessive dehazing caused by misjudgment of fog in the process of dehazing.Then the adaptive weight fusion was used to integrate the characteristic information of different stages,and the preliminary results of defogging were obtained.After the initial dehazing,the image information is more obvious.Then,the preliminary dehazing results are sent to the fine module,and the fog layer in the preliminary dehazing results is extracted by combining the depth information of the image under different resolutions with the encoding and decoding network structure.Finally,the initial dehazing result is subtracted from the fog layer to get the dehazing result.The experimental results show that the performance of the proposed method is superior to that of the mainstream methods.