Research On Depth Image Restoration Technology For Flexible Automated Sorting System

In an industrial assembly line flexible sorting system,it is usually necessary to use a vision camera to realize the positioning and recognition of objects,so as to realize flexible automatic sorting of items.The vision camera realizes the positioning of the object by capturing the depth image of the sorting scene.However,in some cases,due to the particularity of the materials on the surface of the sorted items,such as transparency,high gloss,surface reflection,etc.,the depth information captured by the vision camera is missing in a large area,which affects the robot arm’s performance and grabbing of objects when sorting objects.In view of these existing problems,this paper proposes the following researches on the restoration of depth images for flexible automatic sorting of items:(1)Aiming at the depth image restoration of non-transparent objects with flexible automatic sorting,this paper proposes an effective method for fast repairing the missing depth of the sorted objects.The method first uses the deep network model to quickly and coarsely locate the missing area,and then uses the boundary seed filling algorithm to accurately fit the location hole boundary.Next,for the identified holes,this paper proposes a depth image restoration method based on boundary contour and unsupervised clustering.This method uses the contour of the hole boundary and its neighborhood information to divide the hole into several smooth sub-holes to fit several sub-holes respectively and perform filling hole.(2)Aiming at the depth image restoration of the flexible automatic sorting of transparent objects,this paper proposes a new network structure based on the existing "real-synthetic" transparent object dataset to achieve estimate depth of the transparent objects in the sorting scene.This method first uses a lightweight semantic segmentation model to perform real-time semantic segmentation of transparent objects in the scene to obtain a semantic segmentation mask,and then uses the semantic segmentation mask to remove the depth information of the transparent object in the original depth scene;Then,the modified depth image and RGB image are used as the input of the network learning model,and the depth features and RGB features are extracted through learning,and multi-scale feature fusion is performed.Finally,the depth restoration results of the transparent objects are output.Experiments show that in the depth image repair task of flexible automatic sorting of items,the algorithm in this paper has achieved better results in repair speed and image quality than similar methods and improved the speed and robustness of sorting.