Research on the construction of automatic invoice identification model based on deep learning
Year: 2025
Type: conference-paper
Abstract: In order to improve the recognition accuracy and robustness of multi-class invoice images under structural complexity and interference conditions, an end-to-end automatic recognition model is constructed by integrating convolutional neural network, bidirectional long and short-term memory network and conditional random field. The joint modeling of spatial relationship and semantic structure of characters is realized by introducing location coding and multi-scale feature extraction mechanism. Based on 48,000 multi-source invoice images, the training set is constructed and validated on three types of invoice samples. The average recognition rates of the model on structured fields are 98.28% (electronic), 97.94% (general) and 96.76% (special), and the accuracy of the semi-structured field “amount” reaches 98.4%, which is 6.3% higher than that of CNN. The accuracy of “amount” in the semi-structured field reaches 98.4%, which is 6.3 percentage points higher than that of CNN, showing strong generalization ability and robustness against interference.
Author: Binglan Meng
Institution: Dalian University of Finance and Economics
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416724909