Single Hazy Image Depth Estimation Combining Perceptual Loss Function

Under the condition of haze,there are many water droplets or dust particles in the air,so that light would be scattered or absorpted in the process of transmission,the image quality obtained under the condition is reduced seriously,the image contrast is low,which affects and limits the depth estimation,target tracking,image recognition and other subsequent computer application research.In the field of depth estimation,haze conditions seriously affect the generation of high-quality depth information,so the research on depth estimation under haze conditions has strong practical significance and application value.To solve the problem that indoor and outdoor depth is difficult to estamated under the haze condition,a single haze image depth estimation method combining perceptual loss function is proposed.Based on the convolutional neural network,a two-scale network model is adopted to extract the gloable haze image depth information firstly,and locally refined by combining with the underlying features.Then,the multi-convolutional kernel up-sampling method is used in the up-sampling stage.Finally,the pixel-level loss function and the perceptual loss function are combined into a new composite loss function to train the network.Experiments are trained,tested and verificated in indoor NYU Depth v2 dataset and outdoor Make3 D dataset,the results show that adding multi-convolution kernel upsampling method and the two-scale network model of compound loss function is able to estimate the depth information of the single haze image,improve the precision and quality,and shorten the training time of the model,which improve applicability and accuracy for the estimation of single haze image.Contribution to the point of this article is mainly has the following work:(1)A two-scale model is proposed to estimate the depth of a single haze image.Compared with the existing depth estimation models,this model firstly extracts a single haze image roughly,and then integrates it into a more detailed depth estimation,so that the depth estimation accuracy is higher and more detailed information can be obtained.(2)The multi-convolutional kernel up-sampling method is proposed.In the up-sampling stage,multiple small convolution cores are used to replace the large convolution cores,respectively,and the feature graphs are calculated,and then the obtained images are fused.Small convolution kernel can speed up the operation speed and skip many zero operations at the same time,thus speeding up the network training and improving the image quality of depth estimation.(3)A new composite loss function is proposed in order to get better effect of depth estimation.The composite loss function combines the pixel-level mean square error(MSE)loss and percteual loss,which trained network as the feature extraction,closer to the real depth image on the vision,reduce the noise of the fog cases,to get more information on the details of the real.