Research And System Implemen Tation Of Monocular Depth Estimati On Algorithm Based On Feature Enhancement

The goal of Monocular Depth Estimation is to recover pixel-by-pixel depth information from a single RGB image.Monocular depth estimation provides a theoretical basis for computer vision and promotes the progress of autonomous driving,augmented reality,intelligent robot,and other applications.In recent years,monocular depth estimation has gradually become a research hotspot in academia.At present,the main problems of monocular image depth estimation are:(1)the original image is limited by RGB structure,resulting in insufficient depth feature expression;(2)Fuzzy edges of objects and discontinuous depth inside objects;(3)The appearance structure misestimation caused by texture deviation.The above research difficulties limit the performance of monocular depth estimation and hinder the application of correlation algorithms in real scenes.In view of the existing problems,this paper focuses on the research of feature enhancement,including the following aspects:1.A monocular depth estimation algorithm based on feature refinement is proposed.To solve the problem that the original image is limited by RGB structure,which leads to insufficient expression of depth features,this paper proposes a monocular depth estimation algorithm based on feature refinement.This algorithm aims to improve the diversity of input images and enhance the expression ability of depth features.First of all,in the image preprocessing stage,this paper introduces a multi-mode RGB-depth fusion module to fuse the Depth map with RGB images,enriching the expression of depth features and reducing the noise of input images to a certain extent.Secondly,in the coding stage,a feature refinement module is designed to improve the feature coding ability by refining features layer by layer.Experiments show that the average accuracy of the proposed method on the open data set is equivalent to that of the optimal algorithm,which indicates the effectiveness of the proposed algorithm,2.A monocular depth estimation algorithm based on depth smoothing is proposed.To solve the problems of fuzzy edges and discontinuous depth inside objects,a monocular depth estimation algorithm based on depth smoothing is proposed in this paper,which is used to enrich the detailed information of edge regions and improve the depth continuity inside objects.Firstly,a depth smoothing module is designed to capture the global context by guiding the lower-order features through the higher-order features,so as to improve the depth continuity in the step-by-step feature optimization.Then,this paper also proposes multilevel edge perception smoothing loss to focus on edge information,so as to better perceive the shape change of the object surface.Experimental results show that the proposed algorithm achieves competitive performance on KITTI public data set and the restored depth map has a good visualization effect in depth continuity and edge clarity..3.A monocular depth estimation algorithm based on multi-scale feature correlation enhancement is proposed.In order to solve the problem of incorrect estimation of appearance structure caused by texture deviation,a monocular depth estimation algorithm based on multi-scale feature correlation enhancement is proposed in this paper.The algorithm combines the strengths of Transformer and CNN.Firstly,the Transformer structure is used to capture multi-scale depth features in the input image to fully explore the local details and global structure information.Then,a multi-scale feature correlation enhancement module was designed by CNN to learn the correlation between global features and local features,so as to further improve the global perception ability of the model.This method is helpful to alleviate the problem of misrepresentation of structural information caused by texture deviation.Experiments on the NYU Depth V2 and SUN RGB-D data sets show that the average accuracy of the proposed algorithm on the NYU Depth V2 data set is better than most of the multi-scale fusion methods in recent years,and the accuracy of the algorithm on the SUN RGB-D data set is increased by 4%on average.At the same time,the best accuracy is obtained.Based on the above research,this paper constructs a complete monocular depth estimation algorithm system,which has efficient depth prediction ability and robust feature enhancement ability,and constructs a feasible depth estimation demonstration system.