A real-time and effective text detection method for multi-scale and fuzzy text

The text in the natural scene can be in various forms, dynamic blur and geometric perspective greatly affect the efficiency of text detection. Given the above situation, a real-time and effective text detection method is proposed to detect the multi-scale and fuzzy text. This method applies a convolutional attention mechanism to the feature extraction backbone to obtain more valuable text feature maps. To fully utilize the precise text location signals of the low-level features, a bottom-up path augmentation is used simultaneously. Besides, a few layers of the Resnet-50 backbone are cancelled to further shorten information communication path for balancing the speed and accuracy of detection. For text detection results, the four vertex coordinate values of the text boxes are regressed with the assistance of CIoU loss and shrinkage of text labels. Our model can effectively process an image in the fastest time of 112 ms and has a higher comprehensive indicator value than the other comparative models in ICDAR 2013, ICDAR 2015, and MSRA-TD500 datasets.