Structured Text Reading System

Our approach towards scene-text detection and recognition uses the known regular expression or structure of the target text to bias the model. This regular expression is inputted as a query along with the image. This query guides the detection and recognition process.

Our model can spot the text instances that conform with the query while ignoring the rest of the non-matching text. This process requires minimal post-processing operations. The queries can target text of different lengths, with spaces, or separated in different lines.

In conventional text detection and recognition approaches, the reading system is trained to localize and read text at arbitrary granularities. This is a consequence of the datasets used to train these networks, which usually include the annotations at word level. When these models have to deal with spaced text or text found at different lines, the output of the model is coupled with post-processing operations, an approach which we consider to be sub-optimal. Instead, our training set includes training instances that can have spaces or text in different lines. During the training phase, the model is fed with the image and the query of the target text, which has to be found in the image.