Research On Depth Estimation Algorithm Based On Fast Multi-scale Attention Mechanism

Depth estimation is an important research topic in the field of computer vision.Obtaining depth information is conducive to processing scene analysis tasks.With the development of deep learning,many researchers use methods based on supervised learning to obtain image depth information.However,this method has the problem of lack of real data labels and has certain limitations.Therefore,recent research efforts attempt to use an unsupervised learning strategy to solve the depth estimation problem.Under this research background,in the current depth estimation problem based on unsupervised learning,there are problems that the level of the predicted depth map is not smooth,and the local edge information of the object is blurred,resulting in insufficient prediction accuracy.Innovatively propose a depth estimation algorithm based on a fast multi-scale attention mechanism.The main research contents are as follows:(1)Aiming at the problem that real image data is insufficiently labeled and difficult to obtain,based on the unsupervised learning frameworkthis,introducing attention in Dapth Net,paper adopts a strategy based on unsupervised learning for depth estimation.The framework consists of two parts: the depth estimation network(Dapth Net)and the pose estimation network(Pose Net).The video sequence is used as the input of the depth estimation network,the view synthesis is used as the supervision signal of the network,and the camera motion is combined to estimate the depth information of the image.Due to the interference of noise information in the predicted depth map,a loss function composed of reconstruction loss and smoothing loss is designed.Among them,the reconstruction loss is mainly used to calculate the similarity error between the synthetic view and the real image.Finally,by adding a smoothing loss function,the noise information in the prediction image is effectively eliminated,and a smoother prediction depth image is obtained.(2)Aiming at the problems of low prediction accuracy and blurred edges in the existing depth estimation algorithms,based on the unsupervised learning framework,introducing the attention mechanism in Dapth Net,an image depth estimation model(Fast Multi-scale Attention Model,FSAM)based on a fast multi-scale attention mechanism is proposed.Using an end-to-end Encoder-Decoder structure,the encoder part is used to extract feature information of multiple different scales of the image.In order to capture the remote context information between images and obtain more detailed features in the image,the attention model(AM)is introduced in the network decoder part to effectively improve the contrast between different types of objects,under the premise of ensuring computational efficiency,thereby improving the depth estimation.The problem of inaccurate prediction details and blurred edges.In order to verify the effectiveness of the algorithm in this article,the modules mentioned in this article are trained and tested on the autonomous driving outdoor dataset KITTI and its sub-dataset Eigen Split.Compared with some current classic algorithms in the seven evaluation indicators of logarithmic error,root mean square error,and accuracy under different thresholds,improved accuracy of network prediction.