A document image anti-warping method based on deep learning key point positioning

Abstract

The application discloses a document image anti-warping method based on key point positioning of deep learning, relates to the technical field of digital image recovery, and comprises the following steps: S1, data preparation; S2, model construction; S3, model training; and S4, anti-warping inference. Training data is generated through 3D rendering, a light-weight deep learning network (DenseNet+Transformer) is constructed, the anti-warping problem is converted into key point positioning, and a comprehensive loss function is used to optimize the model. The method does not require additional hardware, the model is efficient, the digital anti-warping of a document picture can be realized at a relatively fast speed and a relatively high accuracy without the aid of any auxiliary photography hardware and 3D reconstruction algorithm, and the method is suitable for large-scale document digitization processing.

Images