Duration: 1:19:34

38,351 views

382

Text in natural images possesses rich information for image understanding. Detecting and recognizing text facilitates many important applications. From a computer vision perspective, text is a structured object made of characters arranged in a line or curve. The unique characteristics of text makes its detection and recognition problems different than that of general objects. In the first part of this talk, I will introduce our recent work on text detection, where we decompose long text into smaller segments and the links between them. A fully-convolutional neural network model is proposed to detect both segments and links at different scales in a single forward pass. In the second part, I will introduce our work on text recognition, where we tackle the structural recognition problem with an end-to-end neural network that outputs character sequences from image pixels. We further incorporate a learnable spatial transformer into this network, in order to handle text of irregular shape with robustness. 

See more at https://www.microsoft.com/en-us/research/video/detecting-and-recognizing-text-in-natural-images/