Researches On Occlusion Edges Extraction Algorithms For Image Depth Ordering Inference

In the field of computer vision,depth ordering inference of monocular images is an algorithm for hierarchical representation of objects in image.Since the depth ordering inference of monocular images can not only get the relative depth between objects in the image,but also restore the accurate contour of objects,and is also the basis of image analysis and image understanding,it plays an important role in object recognition,motion analysis,obstacle avoidance of robot and automatic driving.Generally,this kind of algorithm first extracts the over-segmentation edge of the image through the over-segmentation algorithm of the image,then the contour information of occluded objects is obtained by the algorithm of occlusion edges extraction.Finally,based on the local features of the closed region extracted from the occlusion edges,global depth ordering inference is carried out to obtain the final layer relationship of the object regions.Based on the characteristics of the depth ordering inference process,this paper proposes an occlusion edges extraction algorithm for monocular images depth ordering inference.Firstly,a more accurate superpixel segmentation algorithm is proposed to enhance the accuracy of occlusion edges training samples;then an adaptive parametered classifier is proposed to extract occlusion edges;in the aspect of depth ordering inference,a global inference algorithm based on graph model is proposed to infer the depth order of the separated object regions.The main work of this paper lie in three aspects:superpixel segmentation,occlusion edges extraction and depth ordering inference.(1)superpixel segmentation.Since accurate superpixel segmentation algorithm is the premise and basis of occlusion edges extraction,this paper first proposes a more accurate superpixel segmentation algorithm for edge extraction to enhance the accuracy of occlusion edges training samples.For this reason,this paper combines many edge enhancement features such as color,compactness,entropy and edge terms,exploring the influence of different edge enhancement features on boundary recall of superpixel edges,and proposes a super-pixel segmentation algorithm for multi-edge enhancement features.Then,quantitative and qualitative experiments are carried out on BSDS500 dataset.The experimental results show that the proposed algorithm achieves better edge extraction effect while guaranteeing a promising speed.(2)Occlusion edges extraction.In view of the inaccurate training of the existing Adaboost classifier due to the margin problem,an adaptive parametered Gentle Adaboost classifier is proposed in this paper.This classifier can adjust the weights of the classifier adaptively from each iteration,avoiding the incorrect classification of the samples which is correctly classified in the previous iteration,thus making the training of the classifier more accurate and faster.In addition,because of the testing error of the classifier,there are discontinuity points in occlusion edges.In this paper,the shortest path algorithm and expansion-corrosion algorithm are utilized to optimize occlusion edges.Then the closed occlusion edges is obtained correctly,which can provide an important basis for depth ordering inference.(3)Depth ordering inference.When using the extracted occlusion edges,the T-junction and the convex of edge features are probably have wrong inference of local depth relationship.In this paper,a quadruple descriptor is defined to solve this problem,which can effectively utilize the T-junction and the convex of edge features to correctly judge the depth relationship among different objects.In addition,in order to solve the problem that it is hard to infer the relationship between two separated regions in the existing methods,this paper proposes a new global depth ordering inference algorithm by introducing the graph model,which utilizes the location feature of two separated regions in front view images and the quadruple descriptor.The experiments are done on the Cornell depth order dataset,BSDS500 dataset and NYU2 dataset and got promising results.In this paper,a more effective segmentation and occlusion edges extraction algorithm are proposed,which provide accurate occlude information for depth ordering inference,and finally achieve the depth ordering inference of the whole image.The work of this paper provides important technical support for visual problems such as image understanding and image analysis.