Get The Depth Image From Single View Image With GAN

Depth images indicates the depth information of the object in the scene.The depth information in the scene plays an important role in various visual tasks such as target recognition,scene recognition,semantic segmentation,object tracking,and scene understanding.The generation of the conventional depth image acquires the depth information of the object in the scene by the method of transmitting and receiving the probe light reflected by the object,such as a TOF camera,a Kinect camera that uses structured light to obtain depth information,and a method of obtaining depth information in a scene using lidar scanning.However,the acquisition of traditional depth pictures is greatly affected by the equipment,and the acquisition is difficult.Due to the influence of specific objects in the scene,some depth information may be missing.In the use of neural networks to generate depth images,there are also many studies at home and abroad.The methods are mainly divided into two types,namely,generating a depth picture and a binocular picture through a monocular picture to generate a depth picture.The depth information in the scene can be easily obtained by matching the binocular pictures.Therefore,there is much research on generating depth pictures by using binocular pictures,and there is little research on generating depth pictures by using monocular pictures.A GAN model is proposed to solve the problem of generating depth pictures from monocular pictures.In the generation model,the depth feature is extracted by the structure of convolutional layer and convolutional layer downsampling,and then pixel shuffle is used for upsampling and volume.In the discriminant model,the input of the discriminant network is classified by the convolutional layer and the downsampling structure,and the convolution layer in the FCN is adopted at the end of the discriminant network.Because of the sigmoid activation function,the out of the discriminant model is a two-dimensional probability matrix,so a two-dimensional label is proposed to replace the one-dimensional label in the GAN network,so that the network can achieve the end-to-end training,and finally the results show that the out of the net is good.