Text Spotting for Curved Metal Surface: Clustering, Fitting, and Rectifying

Abstract

In the machining system, how to recognize text on the curved metal surface of parts is still a challenging question. In this article, a text spotting framework for curved metal surface based on image rectification and CNN is proposed. First, a curved text rectification method through CNN-based detection and text curve fitting is investigated. The CNN model for text detection is trained on the original images, which is used to obtain the location of characters of the text. The characters are clustered based on location and size, and the text line is fitted based on the centers of characters that belong to the same cluster. The text image is expanded according to the fitting text line, which finishes the rectification of the curve text image. Second, plenty of original text images are rectified by the proposed rectification method to construct the training data set. Third, in the recognition stage, an original text image is rectified and used as the input of the CNN recognition model. In the experiments, after comparing the detection performance between YOLOv3, Faster R-CNN, SSD-ResNet101, and RetinaNet101, RetinaNet101 was chosen as the text detector. Also, the recognition results on the rectification training set and test set generated by the proposed framework are better than those on the original image data sets. The experimental results show that the proposed method can achieve competitive results text spotting for curved metal surface.