Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (12): 2132-2139.DOI: 10.3778/j.issn.1673-9418.2001021

Previous Articles     Next Articles

Improved Fast Generation of Superpixel Algorithms with Deep Network

SHENG Jiachuan, WANG Jiayuan, LI Yuzhi, WANG Jun   

  1. 1. School of Science and Technology, Tianjin University of Finance & Economics, Tianjin 300222, China
    2. School of Management Science and Engineering, Tianjin University of Finance & Economics, Tianjin 300222, China
  • Online:2020-12-01 Published:2020-12-11

融合深度网络的改进快速生成超像素算法

盛家川王佳媛李玉芝王君   

  1. 1. 天津财经大学 理工学院,天津 300222
    2. 天津财经大学 管理科学与工程学院,天津 300222

Abstract:

Superpixels are the result of over-segmentation of the image and provide an intermediate representation of the image data. It plays an important role in the research of computer vision and other fields. However, the existing superpixel algorithms are non-differentiable and the deep networks are usually defined over regular grid. Thus, most of the current superpixel generation algorithms are based on hand-crafted pixel features. To this end, this paper proposes an improved fast generation of superpixel algorithms with deep network. Firstly, the deep network with multiple hidden layers is directly embedded in the process of superpixel generation to extract the pixel features of the image. Secondly, this paper calculates the initial seed point position by [K-means] clustering method to signi-ficantly improve the segmentation result, and an active search method is used to ensure the correctness of the pixel label. Finally, superpixel segmentation results are obtained. On the Berkeley dataset BSDS500, the benchmark eva-luation using BSDS is compared with other literatures, and the experimental results show that the proposed algo-rithm performs relatively well in terms of the compactness and regularity of segmentation results.

Key words: superpixel segmentation, deep network, clustering, feature extraction

摘要:

超像素是图像过度分割的结果,提供了图像数据的中间级表示,对计算机视觉等领域的研究具有重要意义。现有的超像素算法是不可微的,且深度网络通常在规则的网格上进行定义,导致目前生成超像素的算法大多基于手工提取的像素特征进行。提出融合深度网络的改进快速生成超像素算法,将深度学习网络嵌入到超像素的生成过程中,首先利用含多隐含层的深度网络进行图像像素特征的提取,然后通过K-means聚类方法计算初始种子点位置以改善分割结果,在此基础上通过主动搜索方法确保像素标签的正确性,最后得到超像素分割结果。在Berkeley数据集BSDS500上,使用BSDS的基准测评与其他文献的对比实验表明,所提出的算法在分割结果的紧凑性、规则性等性能方面相对较好。

关键词: 超像素分割, 深度网络, 聚类, 特征提取