Automatic Removal of WaterMarks Logos

Based on machine learning methods, models for automatic search and removal of the logo (WaterMark, waterprint) are developed.The program consists of 3 parts:
Part 1 .Detection, determination of location and cutting the logo with subpixel accuracy.Part 2 .Based on 50-100+ copies of the photo (learning objects) is calculated the reverse mathematical conversion for each pixel, in which the frequency spectrum of the area under the logo corresponds to the area around the logo.This reduces the visibility of the watermark by 90-99% on the original photo.
Part 3 .Package application of the trained watermark removal model for photographs.The Results:
In this work, the logo was removed on 7433 photos.Time of conversion: 3-5 seconds for the cleaning of one photo (1600x1200, 1280x960).The average quadratic deviation is 0.3 gradients of color.Visual control > 93% of photos to recognize the presence of the logo is not possible.6% photo - only with a 300% contrast increase the logo is slightly visible
1% photo - when increased on the site of the remote watermark, a small pixel noise is noticed.Price 200-300$ - training model 1 type of logos
(depending on the complexity of the logo)
N x 0.02$ - Automatic processing of 1st photo

(Examples will be submitted on request)