DetReco: Object-Text Detection and Recognition Based on Deep Neural Network

Figure 1

Overall architecture of the proposed network architecture. Feature maps are firstly extracted with convolutional layers. The object-text detection module is built on top of the feature maps to predict the bounding boxes of the texts and general objects. The NMS module is used to remove the reductant bounding boxes and retain the final positive bounding boxes. The text extractor extracts the text regions corresponding to the coordinates of the text bounding boxes which are the output of the object-text detection module. The text regions are then fed into the text recognition module.