| Popis: |
Under the trend of electronic invoicing recording, replacing labour with automatic key information recognition will help regulators improve their work efficiency. Due to the low image quality of some electronic invoices, the existing text recognition methods still have some drawbacks. On the one hand, some factors such as lighting, tilt, scale, etc., may affect the definition of the photograph in the process of electronic invoicing, resulted in blurring the key information and increasing the difficulty of text recognition. On the other hand, the templates of electronic invoices are often fixed, and the existing methods lack the use of prior knowledge, which increases the computational cost. Given the above facts, this paper proposes a novel knowledge-based key information detection and recognition method for electronic invoices in complex scenes. In the training stage, we simulate complex scenes by using data augmentation techniques such as adding random noise, colour jitter, horizontal lines, and random rotation to improve the accuracy of text recognition. In the modelling stage, the proposed method makes full use of prior knowledge to obtain key information slices to improve the efficiency of recognition processing. Finally, in practice, the effectiveness of the method proposed in this paper is proved. |
