复杂系统管理与控制国家重点实验室知识库

Detecting and recognizing objects in images with complex backgrounds and deformations is a challenging task. In this work, we propose FrameNet, while a deep table lines segmentation network based on our Res18UNet with an adaptive deformation correction algorithm for correcting the table lines. We use Itinerary/Receipt of E-ticket for Air Transport to evaluate our methods. The experiment results show that our Res18UNet can reduce the number of parameters and improve the speed of image segmentation without significantly reducing the segmentation accuracy, and our correction method can better correct the perspective deformation and some distorted tablular images with no dependence on pixel-level ground truth image. In addition, we also apply our model and method to VAT invoice dataset and prove that they also have better transfer ability