Combined learning method for foreground region marking and depth order inferring

Abstract

The invention discloses a combined learning method for foreground region marking and depth order inferring, including the following steps: conducting image segmentation, building image models for the segmented images, building a combined framework based on the predictions of a variety of three tuples, and carrying out training with maximal amplitude on the basis of a structured support vector machine so as to obtain the deep order relationship of each region in the images and the foreground and background markings; The invention makes modifications and improvements to the limitations and deficiencies from one signal predication by deeply estimating monocular images and reviewing and summarizing the foreground region markings, puts forward an estimated combination of framework and makes effective and correct use of some graph verification code algorithms from Geometric Context data set and Cornell Depth Order data set. And a more effective prediction result can be achieved through the method.

Description

technical field [0001] The invention relates to the field of image processing, in particular to a joint learning method of foreground area labeling and depth sequence reasoning. Background technique [0002] The study of depth estimation has always been a fundamental and important issue in the field of computer vision. Researchers initially focused on exploring the absolute depth order of images. Later, due to the difficulty of accurate estimation, they found that the relative depth between objects can be inferred by effectively extracting depth clues of monocular images, such as occlusion and geometric information. sequence, and are used to tackle high-level vision problems like saliency detection and 3D scene understanding. [0003] Most existing techniques use various local cues extracted from contours and T-corner structures to compute relative depth orders. However, this method is naturally flawed. For example, the background area such as the sky and the ground is obv...

AI Technical Summary

Problems solved by technology

If the deviation of the geometric confidence is large, it will seriously affect the final effect, making the system lack of strong robustness

Depth order estimation and foreground area marking are complementary, but existing technical methods do not take this into account, making their connection impossible to be reflected

Images