Research On Depth Image Restoration Method For Intelligent Workshop

In recent years,with the progress of optical technology,3D vision has become an important part of building intelligent workshop,and depth image as an important way to describe the depth information,its quality is very important for 3D vision system.However,due to the limitations of hardware,scene and other conditions,depth images obtained by RGB-D camera usually have holes of different sizes,and incomplete depth information will affect its application.Therefore,in order to solve the problem of depth image holes in the intelligent workshop scene,this paper firstly proposes an intelligent optimization algorithm for image multi-threshold segmentation,and then proposes a depth image inpainting algorithm based on multi-threshold segmentation image guidance.The main research contents are as follows:(1)Research on intelligent optimization algorithm: A new intelligent optimization algorithm,Source Tracing optimization algorithm(ST),is proposed after analyzing the essence of image multi-threshold segmentation and intelligent optimization algorithm.The algorithm emphasizes the information interaction between exploitation set and exploration set.For the two situations when the optimization algorithm falls into the local optimal region,the exploration set uses the historical information of the exploitation set to formulate exploration strategies,respectively.Meanwhile,the exploration set can turn to exploitation set adaptively so as to enhance the exploitation ability.The exploitation set is updated alternately in three update modes,which realizes the simultaneous enhancement of exploitation ability and exploration ability,and improves the convergence speed.In addition,a precocity judgment rule is introduced to ensure timely escape of the algorithm when it falls into the local optimal region.This paper firstly proves the excellent performance of ST algorithm in function optimization problems by using the benchmark function.Secondly,kapur entropy is selected as the optimal objective function,multi-threshold segmentation is carried out for different types of images in BSDS500 dataset according to three different threshold numbers,and compared with several well-known algorithms.The effectiveness of ST in solving image multi-threshold segmentation problem is proved from two aspects of objective data and visual effect.Finally,the algorithm comparison experiment is carried out in the NYU-Depth V2 dataset,which proves that the ST algorithm in this dataset has the best segmentation effect,which provides the basis for the subsequent guided depth image inpainting.(2)Research on depth image repair algorithm: This paper proposes a depth image inpainting algorithm for intelligent workshop scene.Firstly,a filling priority estimation method based on pixel information distribution is proposed.Secondly,a collaborative prediction method combining the exemplar-based inpainting idea and the grey prediction idea is proposed.On the one hand,the exemplar-based inpainting strategy guided by the multi-threshold segmentation image is used for the edge region,and the optimal matching block screening rules and search range are adaptively changed to improve the accuracy and speed of the repair.On the other hand,the grey prediction idea is introduced to repair the smoothing region,and the degree of fit check is added,which solves the problem of the loss of depth gradient information in the traditional method of image repair.The proposed algorithm in this paper is compared with other algorithms in the NYU-Depth V2 dataset.Firstly,the experiment shows that using multi-threshold segmentation image guidance strategy can improve the inpainting speed without affecting the inpainting quality.The average inpainting time is 86.53% of that without guidance strategy,while the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)remain almost unchanged.Secondly,the filling priority estimation,collaborative prediction,and grey prediction idea are evaluated in the ablation experiment separately,and the validity of the innovation points in this paper is verified.Finally,the algorithm in this paper is compared with 5image inpainting algorithms.In all the test images,the average PSNR and SSIM of the algorithm in this paper are 41.732 d B and 0.9862 respectively,while the highest PSNR and SSIM of the other algorithms are 40.019 d B and 0.9816 respectively,the lowest values are only 37.705 d B and 0.9747 respectively.The experimental results demonstrate that our algorithm has its priority to robustness and precision.(3)Intelligent workshop simulation scene depth image restoration: In this paper,Kinect V1 camera is used to capture the image of intelligent workshop simulation scene,including patrol robot,large mechanical arm,small mechanical arm and gas storage tank.The algorithm in this paper is used to repair each depth image,and the repair results prove the accuracy and practicability of the algorithm in inpainting the depth image of intelligent workshop.