Supervised mid-level features for word image representation

摘要

Disclosed is a method and system to learning mid-level features for text images that leverages character bounding box annotations. According to an exemplary embodiment, the disclosed method and system includes extracting semantic local descriptors by aggregating local statistics of small patches and correlating them with character bounding box annotations.

主权项

1. A computer-implemented method for generating a visual feature to semantic space transformation map using a set of text images, each text image including one or more annotated character bounding boxes, the method comprising:
a) extracting a plurality of image patch descriptors representative of a plurality of respective image patches representative of the text images, the plurality of image patches including a background and a foreground area associated with the text images; b) computing a plurality of aggregated representations of the image patch descriptors, each aggregated representation associated with an image block including two or more image patches; c) computing character annotations associated with each image block by measuring a proximate relationship of an image block with an annotated character bounding box; and d) generating the visual feature to semantic space transformation map by constructing an intermediate subspace which maps the aggregated representations associated with the image blocks computed in step b) to the character annotations associated with each image block computer in step c).