Determining the most representative image on a Web page

Abstract

We investigate how to determine the most representative image on a Web page. This problem has not been thoroughly investigated and, up to today, only expert-based algorithms have been proposed in the literature. We attempt to improve the performance of known algorithms with the use of Support Vector Machines (SVM). Besides, our algorithm distinguishes itself from existing literature with the introduction of novel image features, including previously unused meta-data protocols. Also, we design and attempt a less-restrictive ranking methodology in the image preprocessing stage of our algorithm. We find that the application of the SVM framework with our improved classification methodology increases the F 1 score from 27.2% to 38.5%, as compared to a state-of-the-art method. Introducing novel image features and applying backward feature selection, we find that the F 1 score rises to 40.0%. Lastly, we use a class-weighted SVM in order to resolve the imbalance in number of representative images. This final modification improves the classification performance of our algorithm even further to 43.9%, outperforming our benchmark algorithms, including those of Facebook and Google. Suggested beneficiaries are the search engine community, image retrieval community, including the commercial sector due to superior performance.

Automatically Discovering Relevant Images From Web Pages

“…SVM is dramatically the second-worst machine learning method, although it has been suggested in the literature. Considering that Vyas and Frasincar [9] reaches an F-Measure of 0.439 to find the representative image in a different dataset, our performance results are very similar for a different dataset and task. As seen in Table,SVM is insufficient for this task.…”

“…Hu and Bagga [21], Tong and Chang [29], Zang et al [30] use SVM in order to find informative images based on textual queries from users. Vyas and Frasincar [9] focus on determining the most representative image on a web page. They also use SVM and new features to improve their f-Measure score.…”

“…In this task, the researchers mostly concentrate on obtaining textual contents including title [1], [2], main content [3], [5], summary [5], and reviews [6], [7]. Also, there are few studies [8], [9] that focus on automatically extracting relevant images. Moreover, the performance results of these studies are low.…”

“…Previous studies [8], [9] consist of around 1100 web pages which are standard news. However, a website consists of many different pages including standard news, gallery pages, link pages.…”

“…Increasing the number of samples is beneficial for the prediction accuracy of machine learning models. For determining the most representative image on a web page, Vyas and Frasincar [9] construct their prediction model on the machine learning method, Support Vector Machines (SVM). Nonetheless, these studies are not contributive in terms of the comparison of various classification methods.…”