dc.description.abstract | The application of text detection and recognition on optical images is quite extensive. For example, recognition of production date, product part number and drug number, etc... To recognize the text on an image, one has to first detect the bounding box of the text, and then perform the text recognition for the localized image.
However, in order to get a very accurate and robust results under deep learning method, huge amount of data is indispensable for the training of the network model. In addition, before training and testing a deep learning model, it is important to preprocess the image, such as image cropping, scaling and rotating… etc. Data augmentation, which is an approach to increase the number of images, is also important. However, image preprocessing is a very time-consuming and tedious work. In this research, transfer learning is applied to achieve the goal of deep learning training using a small amount of data and get a model with a good accuracy and robustness. In addition to the large amount of data and time required in pre-training a model, the subsequent retrained model can achieve an accuracy higher than 95% in a small amount of text image data.
| en_US |