Research And Application Of Image Occlusion Relationship Reasoning Algorithm Based On Multi-task Network

When projecting a 3D scene onto an image plane,occlusion between objects often occurs,and this occlusion is widely present in 2D images.Occlusion makes the relationship between the objects in the image more complicated,and the information of the occluded objects will be missing,which affects the computer's recognition of the objects in the 2D image and the acquisition and understanding of information.Occlusion relationship reasoning uses algorithms to obtain the occlusion between objects in monocular images and provide valuable object-level depth information,which is a basic problem in the field of computer vision.Occlusion relationship reasoning technology can be used as a key step in other computer vision tasks,and is an important role in scene understanding,3D reconstruction,robot obstacle avoidance and other computer vision fields.At present,with the development of deep learning based on convolutional neural networks,with the help of the powerful feature acquisition and processing capabilities of convolutional neural networks,the effect of occlusion relation reasoning has been significantly improved.The occlusion relationship reasoning based on neural network is divided into two subtasks:occlusion boundary detection and foreground/background relation reasoning.The current popular occlusion relationship reasoning based on convolutional neural networks has begun to learn and predict two tasks at the same time in the same network,but there are still problems of inefficient sharing of network information and feature utilization,and poor extraction of shared occlusion representations.They ignore the effect of joint learning of two or more subtasks on the effect of the occlusion relationship,and also ignore the importance of the representation and prediction methods of foreground/background relation reasoning.And in the application of occlusion relation inference based on convolutional neural network,because it needs the support of professional graphics card with strong computing power,it is difficult to deploy and use in the terminal,which limits the landing and application of occlusion relation inference.Without solving these problems,this article focuses on related work,the main innovative research results have three parts.(1)In terms of occlusion boundary detection,a U-shaped network structure and a trapezoidal network structure that can take into account occlusion and multi-scale are designed,and an attention loss function near the boundary is proposed to deal with the extreme analogy of boundary pixels Balance problem.(2)In the reasoning of the foreground/background relations,the orthogonal representation method of the foreground/background relations,the orthogonal occlusion loss function,and the orthogonal occlusion module suitable for the convolutional neural network are proposed,so that the network can obtain better occlusion information.(3)In the multi-task joint occlusion relationship reasoning,a reasonable semi-shared multi-task joint network structure is designed and a multi-task joint training method based on dynamic sampling is proposed,which improves the performance and efficiency of the reasoning.The algorithm proposed in this paper has been signi ficantly improved in many indicators and visualization results.Compared the F-1 score of fixed thresholds for all images in the datasets(ODS)on PIOD dataset and BSDS ownship dataset to the best method OFNet,our methods is up about 3.9%and 1.6%.And the rest of the metrics have improved to varying degrees compared to OFNet.Finally,based on the research results related to the occlusion relationship inference of the convolutional neural network,this paper designs and implements a cloud occlusion relationship inference system to reduce the terminal's dependence on hardware performance,so as to effectively use computing resources and facilitate deployment.