一种基于像素视觉共生的图像分割方法

A New Image Segmentation Method Based on Pixel-level Visual Co-occurrence

图像分割是图像处理、识别和标注等领域中的重要研究方向。文章提出了一种基于像素视觉共生的图像分割方法,通过对图像中的每个像素提取视觉特征,利用隐层狄利克雷模型进行建模,最后利用像素之间的共生关系获得主题概率并进行分割。与传统的图像分割方法相比,文中所提出的方法能够将图像中不相邻的像素分割为同一类别。实验结果表明,该方法在不同的数据集上能够获得较好的分割效果。

Image segmentation is an important research field of image processing,image annotation and recognition.In this paper,we propose novel image segmentation based on pixel-level visual co-occurrence.First we extract visual features from pixels and then quantized into different visual words.After that,we employ Latent Dirichlet Allocation to model the co-occurrences between different pixels and utilize the generated topic probabilities to segment image into different regions.Compared with traditional segmentation method,the proposed method is able to group non-adjacent pixels into the same semantic region.Experimental results show that our method achieves good segmentation results on different datasets.