Abstract: Aiming at the problem of uneven distribution and  asymmetric characteristics  in practical image distortion. In order to improve the quality of the image and make it closer to the real situation, the distortion in different regions and directions of the image was finely adjusted to restore the original shape of the image, we proposed  a machine learning based asymmetric geometric correction method for distorted images. Firstly, the visual effect of the image was improved and details were enriched by using histogram equalization for brightness compensation. Secondly, we selected some key points or feature points from the preprocessed distorted image, and used the normalized product correlation algorithm to locate and correct all distortion control points required by the positional relationship of these points. Finally, we used the BP neural network in machine learning to learn and fit the complex nonlinear relationship between the original image and the distorted image. Through training, we enabled BP neural network to more accurately describe the distortion characteristics of the image. The network outputs coordinates closed to the control point, thereby achieving asymmetric geometric correction of the distorted image. The experimental results show that the proposed method has good generalization ability and the ability to handle complex asymmetric distortions, which  can effectively improve the accuracy of image distortion correction and increase the average resolution of each image by 465.3 PPI.

Key words: histogram equalization, normalized product correlation algorithm, distortion control point, BP neural network, asymmetric geometric correction