Research On Deep Completion Algorithm Based On Convolutional Neural Network

Depth information can be widely used in important fields such as automatic driving,human-computer interaction,mobile robots,and three-dimensional environment reconstruction.Therefore,obtaining high-quality depth perception information from the scene is an important research direction in the field of computer vision.However,the original depth image obtained by the depth sensor has problems such as sparseness and missing pixels in some areas,and the reliability is not high,and it is difficult to directly apply to practical tasks.In order to meet the actual needs and obtain high-quality and dense depth images,in addition to improving the hardware equipment,researchers have conducted a long period of research in the field of depth image completion from the algorithm level.Recently,with the development of convolutional neural network research,due to its powerful representation ability,the academic community has proposed many deep completion algorithms based on convolutional neural network.However,it is still a challenge to save computing resources as much as possible while reducing the amount of network parameters while obtaining excellent completion results.This paper mainly analyzes the development context of the depth completion algorithm,with light weight and high performance as the research goal.Aiming at the defects of the original depth image and the deficiencies of the existing depth completion algorithm,a dual-branch,multi-stage,multi-scale color image guide is proposed.The depth image completion network.Specifically,this article discusses and studies a dual-branch,multi-stage,multi-scale color image-guided depth image completion network.The algorithm includes color image-guided branches and depth-refined completion branches,using color images and defects in the same scene.The depth image is used as input,and the coarse completion result is obtained through the color image-guided reconstruction branch,and then the depth image refinement and completion branch is used to obtain the refined and completed depth image.Among them,the feature images of the four resolutions in the decoding end of the color image guiding branch are spliced to the decoding end of the corresponding resolution of the depth refinement completion branch to achieve the purpose of multi-scale fusion.Design a multi-scale convolution with random channel shuffling to enhance the feature extraction ability of the network and achieve the purpose of improving the completion effect.The channel perception mechanism is used to filter the feature images that are beneficial to the depth completion task in the multi-scale fusion process and the refined reconstruction process.Finally,the use of multi-stage loss function technology further improves the effect of depth completion.Finally,this paper conducts a lot of experiments on actual indoor and outdoor data sets.Through experiments,a qualitative and quantitative comparison between the deep completion network proposed in this paper and the existing mainstream deep completion algorithms has verified the superiority of the deep completion network proposed in this paper and the rationality of its structural design.