基于心理量表的情感图像检索性能评价方法研究

Study on evaluation methods of emotion image retrieval using psychological scale

情感图像检索中检索算法优劣的有效评估是一个尚未解决的问题。本文提出基于心理量表的情感图像检索性能评价方法,采用等级排列法或对偶比较法对不同算法检索出的图像进行排序,建立顺序量表,并在样本满足正态分布假设的前提下,将顺序量表转化成等距量表,以比较检索算法的优劣。文中将本实验室实现的情感检索算法检索出的"厚重的"、"简洁的"服装图像以及"热情的"、"视野宽广的"风景图像与随机检索出的图像进行了比较,验证了基于心理量表的情感图像检索性能评价方法的有效性。

How to evaluate the effectiveness of emotion image retrieval algorithms is an open question. This paper presents an evaluation method based on psychological scale. Rank-order method or method of paired comparison is adopted to sort the images retrieved by different algorithms, then an ordinal scale is obtained. With presupposition of normal distribution, an equal interval scale is also obtained though POZ conversion. Thus different algorithms can be compared on these scales. We experimentally evaluate the effect of the proposed method using our emotion retrieval algorithm and random algorithm. Four kinds of emotion images, thick fashion images, succinct fashion images, passional scenery images and wide view scenery images are retrieved and compared. The experimental results show that our retrieval algorithm is better than random algorithm