Superpixel fusion method of local correlation

Abstract

The present invention provides a superpixel fusion method of local correlation which enables the iteration operand to be reduced and the fusion speed to be accelerated, and belongs to the image segmentation field in the image processing. The superpixel fusion method of the present invention enables the local correlation between the pixels to be enhanced by taking the pixels as the basic fusion units, and further realizes the superpixel fusion by the interaction between the pixels. During the pixel fusion process, the division of the superpixels is used as the prior guidance information, thereby guaranteeing that the pixels in the superpixels have the same categorical attributes, and keeping the edge discrimination capability between the object areas while increasing the correlation between the adjacent pixels. By the superpixel fusion, the semantic feature space of each fusion area is obtained, and the corresponding fusion is carried out according to the correlation. During the single pixel classification process, and by converting into different semantic spaces, the distances from the pixels to the clustering centers are calculated according to different visual focuses, thereby improving the reliability of the pixel categorical attribution. Finally, the superpixel fusion is realized according to the mark of the superpixels under an original resolution.

Description

technical field [0001] The invention relates to a superpixel fusion method, in particular to a locally correlated superpixel fusion method, and belongs to the field of image segmentation in image processing. Background technique [0002] Image segmentation is the first and key step in image analysis, and the result of segmentation directly affects the analysis of subsequent image content. If there is a lack of corresponding background knowledge, it is still a difficult problem in the field of image processing to realize image segmentation only from the visual content itself. In view of the fact that accurate image segmentation is difficult to achieve, only weak segmentation methods can be used in practical applications, and then the weak segmentation results are synthesized in subsequent steps in order to obtain reasonable segmentation results. Superpixels are currently the most commonly used image weak segmentation results, and have become a hot spot in image segmentation....

AI Technical Summary

Problems solved by technology

For the texture area, due to the large visual difference between texture elements, the same texture area may be divided into superpixels with large visual difference, which will bring greater difficulties to the subsequent fusion

Superpixel fusion should not only consider the visual similarity between superpixels, but also consider the positional correlation between superpixels, especially for the fusion of texture regions, the correlation in spatial position is particularly important, and Marko The Markov model is an ideal tool to express this local correlation, so although there are many superpixel fusion methods, they are basically limited to the Markov random field or the related conditional random field model. However, the irregular shape of the superpixel and the coupling between the parameters of the Markov model lead to the high computational complexity of this type of method.

Images