The uploaded photo images by customers cannot be directly used in the following fake recognition procedures, due to the missing of details, the image blurs, the unclear boundary, the low resolution etc. To solve these problem, the low quality photo image to high quality scan image transformation step is the fundamental step for pre-processing the input images, after image-to-scan transfer, we can more easily implement the following functions:

  • Face Recognition
  • Text Matching
  • Barcode Reading

The basic idea for image-to-scan has three major components, 1, Accurately locate segment the ID area of interest; 2, Transform the target area to target size by task-specific affine transformation; 3, Precisely generate the unknown/missing areas by Generative Adversarial Network (GAN).



Both segmentation task and GAN task are based on deep learning model and big data driven. The model is trained with around 5,000 collected/annotated data. The following affine transformation is a geometric transformation that preserves lines and parallelism (but not necessarily distance and angles). The final output image will be a scan-shape image. Compared with regular input photo-shape, the scan-shape has the following advantages:

  • Clearer features on the image
  • Less blurs, clearer boundary
  • Repair the missed/broken parts of input IDs
  • High speed for transformation