Web Scraping of online newspapers via Image Matching

Reading is an activity which takes place widely on the web: every newspaper in the world has his own digital version on the internet and there are even a lot of magazines that exist only on the web. In such a scenario, Computer Vision can offer a useful set of tools that can help web editors to improve the quality of the provided service. One of these tools is here presented: given a webpage of a newspaper or journal, the proposed framework localizes news items remotely clicked by users, giving the bounding box of the content of an article in its relative homepage. The tool is hence able to track an article in the page in which is contained at any time during the day: such an information is very useful for web editors to understand the trend of the published items and to rearrange the contents of the homepage accordingly. The system has been developed with an hybrid approach: first we manipulate the HTML source of the homepage in order to generate a visual template for the news item we want to localize. A visual template for a news item is an HTML document containing only the displayed content of the article in the homepage, that is the text of the title and the image attached to the article, if any. The algorithm which builds the template needs just the URL of the article. Thereafter, we take the screenshot of the template in order to get an image representing the news. On this image, we extract FAST keypoints [1], which fit well for the kind of images we are dealing with, that is almost blank images mainly composed of text. The keypoints extraction is done also for the screenshot of the homepage at system start time: therefore, we can now match the article image with the homepage image and get the bounding box of the news in its parent webpage. We provided four keypoint descriptors for the matching: SIFT [2], SURF [3], BRIEF [4], BRISK [5]. We also implemented articles localization via Template Matching, using the Normalized Correlation Coefficient as measure of similarity. In order to ensure the matching, we have to discard all the redundant blank content placed on the borders of the image, given by the fact the picture comes from a screenshot. We use the extracted keypoints for this purpose: since FAST keypoints are located only upon non blank parts of an image, we can use the bounding box containing them to cut the picture and remove the void borders. We tested the system by labelling the bounding boxes for all articles for a set of websites. Then, we compared the extracted rectangles with the ones of the ground truth. We compared the performances of Template Matching versus Keypoint Matching using the aforementioned descriptors. For each extracted bounding box, we measured the euclidean distance of its center from the corresponding rectangle center of the ground truth; for each test, we provide evaluation measures to assess the accuracy of the developed algorithm. Also, we provide HIT/MISS. The results depend on the layout of the websites and on the manner they arrange the contents in the homepage. Our tests show that good performances can be reached, giving also an interesting case of comparison among different keypoint descriptors.