A method of representing superpixels and measuring the similarity between superpixels

Abstract

The invention discloses a method for expressing ultrapixels and measuring similarity between the ultrapixels. The method comprises two phases. According to the first phase, an ultrapixel BoS is built, namely, the ultrapixels are acquired regarding to a given input image, an ultrapixel pyramid is built, the maximum inscribed square is extracted, outstanding square calculation is carried out, and ultrapixel BoS descriptors are built; according to the second phase, ultrapixel similarity calculation is carried out based on an ultrapixel BoS model, and the similarity of two ultrapixels is measured on the basis that the ultrapixel BoS descriptors are acquired. Due to the adoption of the method, ultrapixel characteristics can be better matched, detailed information of the ultrapixels can be precisely captured, and ultrapixel separation and merging diversity can be reasonably dealt with. The method can ensure that ultrapixel-based specific application algorithms like image cutting algorithms and image matching algorithms in relevant fields like computer visions can better copy with irregularity, separation-merging diversity and multi-characteristic characteristics so as to achieve better effect.

Description

technical field [0001] The invention relates to a method of image processing and image analysis. In particular, it relates to a new method of representing superpixels and measuring the similarity between superpixels applied by superpixel-based algorithms. Background technique [0002] Superpixels are perceptually meaningful irregular regions in an image that are often created by merging similar neighboring pixels. An important issue in superpixel-level applications is how to reliably measure the similarity between two superpixels. Most of the current work related to superpixels mainly focuses on the generation of superpixels, or the application of superpixels in different computer vision problems. Compared with regular image regions, superpixels have the following point features: (1) Irregularity, the superpixels of a picture have different sizes and shapes. (2) Separation-merging diversification, the same object may get different numbers of superpixels with different siz...

AI Technical Summary

Problems solved by technology

(3) Multi-featured, most of the superpixels generated by superpixel algorithms are mostly consistent regions, but due to parameter settings, algorithm failure and other reasons, there will be superpixels containing multiple consistent regions

Even though these methods perform well in some applications, they ignore certain features in superpixels that would invalidate general image region features used when measuring superpixel-level similarity

Images