Authors: Liu Rong, Xu Chengshen, Huang Xiao, Li lin
Description: For this Text localiztion task, we detect the position of the text lines via a deep-learning algorithm based on the most popular object detection network Yolov3. Small amount of data augmentation strategies were applied for the effective utilisation of provided training samples. Two steps are carried out before detecting the individual text regions in order to impove the MAP scores of the localization. First, we cut a rectangular region tightly enclosing the text areas from every original images to excluding the effect of the large whitespace on the edge. Second, images with extremly large height/width ratio will be cut into two images before detecting so that the input images of the yolov3 network own a proper aspect ratio.