Text Spotting on Metal Part Identifier
Li Liu, Wenbo Sun · 2021 3rd International Academic Exchange Conference on Science and Technology Innovation (IAECST) · 2021
In an industrial scene, the text had a more disturbing background and unique font style with depression, hump and other phenomena. And this type of text accounts for a very low proportion of the public datasets which makes the existing models unable to perform well in the detection and recognition of industrial scene text. In view of the small industrial scene text dataset, the dataset was expanded through data augmentation operations to enhance the generalization ability of the model. Based on the Point Gathering Network, the trained model was transferred to the augmented metal part word data set through the Fine-tune method for retraining to improve the model performance. The experimental results show that the F-score of detection is 81.3% and the accuracy of recognition is 86.5%, which basically realizes the detection and recognition of metal parts Identifier.