The Reaserch Of Occlusion Boundary Extraction Algorithm For Image Depth Order Reasoning

Image,as an important information carrier,widely exist in our lives.It is always considered as an important ability of understanding and perception to restore the depth order of objects from a single image.In the field of computer vision,image depth order reasoning is based on the occluded edge in the image,and describes the hierarchical relationship between regions or objects in the image.It is a basic and challenging problem,which is important for image analysis and understanding.It can be used to assist many high-level visual perception tasks,such as object detection/tracking,motion analysis,structural motion and 3D reconstruction,automatic driving,scene understanding,etc.Although people can easily perceive the depth order between objects in the image,it is very difficult for computers to perceive the depth order in the image.By studying human perception,we find that edge location in natural images affects the levels of computer vision perception.The accurate location of occluded edges can effectively help the computer understand and split the high-level perception tasks as human beings do,and better recover the depth order relationship between objects from a single image.Therefore,occlusion edge extraction is undoubtedly the most basic and key step in depth order reasoning.At present,there are two kinds of occlusion boundary extraction algorithms in depth order reasoning:the edge-based occlusion boundary extraction algorithm and the pixel-based occlusion boundary extraction network.We explore and analyze the problems in the edge-based occlusion boundary extraction algorithm and pixel-based occlusion boundary extraction networks.To solve the problems of the existing occlusion boundary methods,we have proposed more accurate occlusion boundary extraction algorithms to optimize the results of depth order reasoning.The main achievements are as follows:(1)An edge-based adaptive occlusion boundary extraction algorithm is proposed and applied into the edge-based depth ordering reasoning algorithm.Firstly,we propose an adaptive super-pixel segmentation algorithm called the adaptive DRW to modify the segmentation adaptivity,boundary adherence and segmentation shape.We improve the quality of segmentation results from three points:the adaptive seed initialization,the adaptive weight decay function and the shape constraint,making the segmentation results more suitable for occlusion feature extraction and occlusion relationship judgment.Secondly,in order to further improve the accuracy of occlusion relationship judgment,we explore and enhance occlusion relationship judgment from the two levels:feature calculation and classification.As for the existing feature calculation,we enrich the kinds of occlusion features and further consider the feature instability factor during feature calculation.To improve the classification of unbalanced data,we propose an adaptive cost-sensitive algorithm--adaptive AdaCost.In the proposed adaptive AdaCost,we design the adaptive cost term,which can dynamically adjust the cumulative importance of samples,and make the classifier pay more attention to the difficult samples of the minor classes.Then,we use the adaptive cost adjustment term to further reduce the cumulative misclassification cost objective function.Eventually,by rescaling the cumulative misclassification cost,we get the upper bound function of the cumulative misclassification cost and utilize the approximation solution of the upper bound function to realize the practical training and using of the adaptive AdaCost,which can better classify the unbalanced occlusion boundary.Finally,based on the proposed adaptive DRW and AdaCost,we construct our edge-based adaptive occlusion boundary extraction algorithm and embed it into the edge-based depth order reasoning algorithm framework to obtain the final depth order.Qualitative and quantitative experimental results show that our adaptive DRW and AdaCost have better self-adaptability,practicability and performance compared with other segmentation and classification algorithms.Besides,our adaptive occlusion boundary extraction algorithm can make the depth order reasoning performance significantly improved,which shows the practicability,effectiveness and innovation of our proposed edge-based adaptive occlusion boundary extraction algorithm.(2)A new pixel-based occlusion extraction network—Mutual Boundary-Orientation Occlusion Network(MBOONet)is proposed and used for pixel-based depth order reasoning.Starting with the current pixel-based occlusion representation,we first redefine the representation of occlusion boundary and analyze the dialectical relationship between boundary and orientation,according to which we simplify the representation of occlusion boundary and re-plan the learning tasks of boundary and orientation.Then,we put forward the Mutual Boundary-Orientation Occlusion Network,which adopts "shrinking-radiation" dense connection mode and the multi loss supervision mechanism with progressive supplement,so as to further improve the accuracy and clarity of boundary and provide more accurate edge guidance for occlusion orientation learning.According to the dialectical relationship between edge and orientation,we take the boundary result as the guidance information to instruct the orientation learning,and recover the boundary again from the orientation result,which ensures the internal consistency of the boundary and orientation in the representation of occlusion boundary and enhance the practicality of occlusion orientation.Finally,we embed it into the algorithm framework of pixel-based depth order reasoning to get the final depth order.We carry out the experimental verification in the quantitative and qualitative ways,by comparing our MBOONet with the advanced occluded boundary extraction network on the PIOD dataset,and the edge detection structure with the edge detection model HED on the BSDS500 dataset.The visual and experimental results further prove the rationality of the proposed occlusion boundary representation,the novelty and effectiveness of the occlusion boundary extraction network model.In this paper,the edge-based and pixel-based occlusion boundary extraction algorithms are studied to provide accurate occlusion boundary for depth order reasoning.The work in this paper provides a feasible solution and technical route for high-level vision problems,such as scene understanding etc.,which need to accurately perceive the edge and relative location of objects.