Depth order reasoning method based on occlusion feature learning

Abstract

The present invention provides a deep order reasoning method based on occlusion feature learning, comprising the following steps: obtaining an over-segmented image through the Berkeley algorithm, using an occlusion edge classifier to extract over-segmented edges from common edges; after color model processing, Extract occlusion edge features; select sparse-based classifiers, train and determine occlusion edges; remove false edges by fusing same-layer regions, and avoid misjudgment problems in the process of occlusion edge classifiers; adopt a new triple description operator to realize semi-local depth order reasoning; finally, through modeling, describe the local depth order as a partial order relationship, and use the relevant knowledge of graph theory to transform the global depth order reasoning into a kind of The problem of effective path, so as to complete the depth order reasoning of the whole image.

Description

technical field [0001] The invention relates to the research and realization of a deep sequence reasoning method based on occlusion feature learning. Background technique [0002] The depth order reasoning of monocular images was originally proposed by N.Mark and D.Mumford in 1990. The depth order reasoning of monocular images includes two processes: the process of segmenting the image into non-overlapping regions and the process of sorting the segmented regions according to the occlusion relationship. Regions are the parts between edges. Therefore, the process of dividing an image into non-overlapping regions can be transformed into a process of detecting edges, especially occluded edges. P.Arbelaez and J.Malik proposed the method of edge detection. However, the method they use is only applicable to hierarchical probability margin maps, and noisy margins can seriously affect the results of depth order inference. In contrast, the method based on occlusion relationship ed...

AI Technical Summary

Problems solved by technology

Depth order inference of some existing monocular images, because of the lack of 3D information in the scene, the effect of depth order inference is not very good

Images