Research on Curved Chinese Document Correction Based on Deep Neural Network

Aiming at the correction of curved Chinese document image, a deep neural network-based parameter estimation method was proposed for estimating document deformation and camera pose, with which the correction could be performed straightforwardly. The normalized coordinates of characters of several consecutive lines were used as input to the network, which will output those parameters. The samples for training and testing were automatically generated based on polynomial curve-model and perspective projection principle, with random noises. Experiments shew that with corrected document images, average OCR rate is 5.701% higher than traditional four-point correction method.