A Multiscale Cascaded Hierarchical Model Approach for Enhanced Image Labeling

Abstract

The invention discloses a multi-scale cascading hierarchical model method for improving an image marking effect, and the method comprises the following steps: 1, an initialization step: building a hierarchical scale space of a data set and enabling a quality matrix of a to-be-marked image to be initialized to be the worst; 2, selecting an image marking method, making an original image, and updating the marking quality matrix of the original image; 3, employing the same marking method to mark other image layers of the scale space, and updating the marking quality matrix of the image, wherein parameters of a marking process are generated according to marking results of the original image; 4, carrying out scale restoration of the marking results of the other layers, and updating the marking results of the original image according to the obtained marking quality of each layer; 5, repeating step 3 and step 4 till the quality matrix of the original image does not change after updating. The method employs that things with different scales in different semantic domains are different in expressive force in the same image, and can effectively improve the marking accuracy.

Description

technical field [0001] The invention belongs to the field of computer vision methods, and relates to a multi-scale cascaded hierarchical model method for enhancing image marking effect Background technique [0002] The task of image labeling is to label each pixel in an image with its semantic category (labeling each pixel in an image with its semantic category), which is an important step and basis for scene understanding and plays an important role in the field of computer vision. In the past research, many effective labeling methods have been proposed, such as template matching, association method, feature bag method, shape model, label transfer method, etc. These methods consider using fixed-scale input images and fixed-scale object categories on the input images, and train discriminative models for each labeled category from some fixed-scale training data. At the analytical level, these systems use learned or matched models formed by pixels, windows, edges, or other im...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the problem of low accuracy of image labeling results in the prior art, the present invention proposes a multi-scale cascaded hierarchical model method for enhancing image labeling effects, using the characteristics of different expressive power of images at different scales and the model Run the convergent iterative algorithm to enhance the single-layer image marking effect, the calculation structure is simple, and the convergence is good

Images