Research On Depth Image Estimation And Enhancement Technologies For Complex 3D Environment

High-quality depth information has been widely used in computer vision tasks such as autonomous driving,robot navigation,3D reconstruction,and object recognition.However,although the existing depth cameras can be used to collect the depth information required for the above tasks,due to the limitations of their hardware devices,their perceived depth distance is limited and the collected depth information has problems such as low resolution and large noise,which makes the depth There are certain limitations in the application of cameras in different scenarios.In view of the above problems,the enhancement of low-resolution and low-quality depth images and the prediction of high-quality monocular depth images have received extensive attention from academia and industry.At the same time,the existing color image-guided depth image super-resolution methods have problems such as high model complexity,a large number of parameters,and slow operation speed.In addition,most deep learning-based methods often lack interpretability of the model,which will seriously affect the performance,application scope,and rationality of the model.In response to the above problems,this paper studies the following parts:(1)A depth map estimation method based on deep learning hybrid cross-guided filtering is proposed.Specifically,the whole network is divided into two parts : rough monocular depth map estimation model and depth map super-resolution reconstruction model.Firstly,a simple U-shaped network is used to obtain a rough low-resolution depth map in the rough monocular depth map estimation stage.Secondly,considering that the reconstruction results of the existing color-guided depth map super-resolution reconstruction method have serious detail loss and structural distortion at extremely high sampling rates,and the existing network has high computational complexity.In addition.The inconsistency between color map and depth map makes these methods unable to fuse bimodal features well.In order to obtain high-quality reconstructed depth maps in the depth map super-resolution reconstruction phase,a hybrid cross-guided filtering model for color-guided depth map super-resolution reconstruction is proposed,which uses the consistency of color depth maps from multiple perspectives to gradually improve the quality of depth maps.A large number of experiments show that the performance of the proposed monocular depth estimation method is better than many of the latest methods.At the same time,the depth map super-resolution reconstruction model of the proposed method is extended and applied to solve the problem of color map guided depth map super-resolution reconstruction.(2)Inspired by multi-task learning,a joint multi-task model-driven low-quality color image-depth image enhancement network is proposed.Specifically,by using maximum a posteriori estimation,the degraded low-quality color image enhancement and depth image enhancement tasks are converted into an optimization model that combines color and depth information.The model is alternately optimized in an iterative manner to complete the color image guided depth map super-resolution task and the low-brightness color image enhancement task.The whole iterative optimization process is extended to a joint multi-task optimization model.A large number of experimental results confirm that using the same network can solve the high-resolution reconstruction of depth maps and the enhancement of low-brightness images in parallel.In addition,the proposed interpretable network method can outperform many uninterpretable color image-guided depth map super-resolution methods and low-brightness color image enhancement methods.(3)A single depth map denoising method is proposed that combines high-low frequency decomposition and multi-scale two-level fusion strategy.Specifically,a multi-scale Gaussian filter is first introduced to decompose the noisy depth map into a set of low-frequency structural components and a set of high-frequency detail components.Secondly,in order to make full use of the complementary characteristics between high and low-frequency information,these two groups of components are respectively input into the low-frequency feature extraction network and high-frequency feature extraction networkl,and the complementary feature weighted fusion mechanism is used to realize multi-level feature Fusion and feedback.Finally,a high and low-frequency combined reconstruction module is constructed to perform residual prediction on the high and low-frequency enhanced features output by the high and low-frequency feature extraction network and add them to the input image pixel by pixel to obtain a high-quality depth image.Experimental results show that this framework has better performance than several mainstream depth image denoising methods in terms of peak signal-to-noise ratio,root mean square error,structural similarity,and comprehensive performance comparison.