Web Table Classification Based on Visual Features

Tables on the web constitute a valuable data source for many applications, like factual search and knowledge base augmentation. However, as genuine tables containing relational only account small proportion of web, reliable table classification is crucial first step extraction. Previous works usually rely explicit feature construction from HTML code. In contrast, we propose an approach by exploiting full visual appearance table, which purely applying convolutional neural network rendered image table. Since these features can be extracted automatically, our circumvents need construction. A new hand labeled gold standard dataset code images 13,112 was generated this task. Transfer learning techniques are applied to well known VGG16 ResNet50 architectures. The evaluation CNN with fine tuned (F1 93.29%) shows that achieves results comparable previous solutions using explicitly defined based features. By combining features, F-measure 93.70% achieved Random Forest classification, beats current state art methods.