摘要：AbstractIn recent years, the affect-based image indexing by visual features is going to be popular research area by increasing the importance of affective computing. So far many algorithms and systems have been developed to index images using affects and objects, then their performance is highly dependent on the training data with accurate labels. Most of the existing systems have generated the ground truth based on manually tagging by human, which is too time-consuming and costly subjective of image. Accordingly, a new mechanism to collect new trustable ground truth is presented, which is named by “Image Battle”, to collect more trustable ground truth. The Image Battle system consists of three modules: image crawling, image voting and image ranking. First, for a text query, the images are first crawled then they are filtered to remove some noisy data. In the second stage, two images are randomly selected from the database and are evaluated by receiving votes from participants. After performing this process about several times over a period of two or three months, the system computes rank values for every images based on their number of wins and losses at the evaluation. These three procedures can be iterated whenever new data are added to the database, to update the ranks of images. To validity the effectiveness of the proposed system, the generated ground truth through image battle are used in some researches to train the affect-based image indexing system. When compared with the existing system, the proposed system can improve the accuracy. In addition, it is proven that the image battle system can provide more perspective and convenient interface to collect users’ evaluations.
关键词：Image Battle System;Affect-based Image indexing;ImageRank;PageRank;Probabilistic Affective Model (PAM)
