Computer Engineering and Applications ›› 2019, Vol. 55 ›› Issue (12): 215-219.DOI: 10.3778/j.issn.1002-8331.1803-0460

Previous Articles     Next Articles

Development and Application of Improved SSIM Image Evaluation System

ZHU Jiong, YAN Lijun, MA Yan   

  1. College of Information and Mechatronics, Shanghai Normal University, Shanghai 200234, China
  • Online:2019-06-15 Published:2019-06-13

改进SSIM图像评价系统的开发与应用

朱  炯,严丽军,马  燕   

  1. 上海师范大学 信息与机电学院,上海 200234

Abstract: In order to realize the automation of image evaluation in all kinds of course and examination, and reflect the degree of difficulty of different content in one image, the image automatic evaluation program is designed. By combining the watershed algorithm with the Structural Similarity(SSIM) theory, a WSSIM algorithm for evaluating the degree of difficulty is proposed. A total of 3,173 samples are tested by randomly selecting from 5 actual experiments, which already have artificial evaluation results. By calculating the absolute error between the artificial evaluation and the system evaluation, the results show that:more than 63.75% of samples error is less than 1.5, while the maximum value can reach 74.26%. More than 84.41% of samples error value is less than 2.5, while the maximum value can reach 92.68%. The statistical results show that the results of system evaluation are in good agreement with manual evaluation, and the standard difference is less than manual evaluation. The average accuracy of WSSIM is 12.91% higher than that of SSIM. The experimental results show that the system is effective and can be run directly in the Windows 7 after compilation. It can help relevant institutions and grade examination to carry out automatic image marking.

Key words: image similarity, Structural Similarity(SSIM), watershed algorithm

摘要: 为了实现图像内容相关评价中的自动化,体现出一幅图像中不同内容的难易度,实现图像评价的智能化,开发了图像自动评价系统。将分水岭算法(Watershed Algorithm)和结构相似度(SSIM)理论相结合,提出了一种针对难易程度进行评价的算法WSSIM。通过随机选取5场实际考试进行检验,共计3?173个样本。计算了人工评价与系统评价绝对误差,结果表明:所有的实验组中均有超过63.75%的样本绝对误差数值小于1.5,最大值达到74.26%,有超过84.41%的样本绝对误差数值小于2.5,最大值达到92.68%。统计结果表明:图像的自动评价结果与人工评价有比较好的一致性,标准差值小于人工评价,改进后的方法能使绝对误差数值小于1.5的样本数量平均提高2.77%。实验结果表明:该系统是有效的,经编译完成后,可在Windows7操作系统中直接运行。它能够为相关院校和等级考试中图像的自动评价提供帮助。

关键词: 图像相似度, 结构相似度, 分水岭算法