Research On Depth Estimation From A Single Image Based On Residual Dense Network

Depth estimation from image has been an important issue in the field of computer vision.It has extremely important applications in intelligent robots,semantic understanding,3D scene reconstruction and autonomous driving.It is the most difficult to obtain depth information from a single image,because different 3D scenes can be projected to the same 2D images,humans can accurately determine the depth information of a single image based on rich prior knowledge,but it is very difficult for the computer.In recent years,neural network has been increasingly used in the field of depth estimation from single image,and researchers have proposed a large number of methods based on neural network.Compared with traditional methods,the method based on neural network can get a more accurate depth estimation,but there is still a big gap between the result and the real depth map.First of all,this paper researches and analyzes the research status of depth estimation from single image at home and abroad,and finds that most of the existing methods of depth estimation from single image can not give a clear depth representation of the edge information of the object in the original image.Therefore,this paper proposes a method of depth estimation from single image based on Residual Dense Network.By introducing Residual Dense Module into the encoder-decoder with skip connection,this paper proposes Residual Dense Network suitable for depth estimation from single image.The self-supervised training of the network is achieved by using binocular stereo image pairs.By introducing the Residual Dense Module,the network can better obtain the local features of the image,and the ability of method to extract the edge information of object can be improve.Secondly,by researching the depth estimation map of the method of depth estimation from single image based on neural network,it is found that most network models have low feature extraction capabilities for areas with small depth gradients in the image,which affects the accuracy of the depth estimation results.Therefore,this paper proposes a method of depth estimation from single image combined with Deep Layer Aggregation.According to the theory of deep aggregation,Deep Layer Aggregation Network is proposed.Deep Layer Aggregation Network is used to improve the extraction capability of high-dimensional features of Residual Dense Network,which makes the network more sensitive to the deep gradient domain,further improving the accuracy of depth estimation for regions with small depth gradient.Finally,KITTI dataset is used to train and test the network in this paper.By comparison,it is found that the method in this paper can get better accuracy and error values than most existing methods.It proves the effectiveness and superiority of this paper in depth estimation from single image.