This project aims to build a software to detect a piece of paper from an image and build a new one with just the paper without any type of mark (pen, coffee...) and with the perspective adjusted. It has been developed using computer vision techniques with Python and Scikit-Image.
To use it, you can choose an image from the Examples folder by the number of the document, or specify a path:
python main.py <number_or_path>The processing is done by solving smaller individual problems, and putting them together to generate a global solution. The following sections describe each of the problems identified and how it has been solved.
This step is not necessary for images with enough contrast between the piece of paper and the background. However, for those ones with a light or even white background it is almost compulsory.
To enhance the contrasts, a gamma adjustement is applied to the image, so the edges of the paper become clearer.
First of all, a downsize is performed. After that, a median filter is applied in order to remove the text so only the important edges are left. Then, the Canny operator and the Hough transformation for lines is applied. We then classify the four strongest lines in vertical and horizontal lines. Finally, the intersection of the ones in the first group with the ones in the second group is find, those are the corners. Extrapolating to the original size of the image, we have the coordinates of each of the corners.
To avoid errors due to the aproximation and the distorsion made by the filters, a little correction is made to the corners so they are a bit inside the paper. After this, the projective transformation is made by considering that each of the corners correspond to each of the corners of the image.
To detect whether a pixel is text/mark or the paper itself, a local thresholding is perform. This get rids of lights, shadows and transparent marks such as coffee droplets.
To differenciate between what is text and what is a mark, the standard deviation of the three RGB components of each pixel that is not background is done. Notice that in this step the text that is tinted by, for example coffee, is marked as color/mark.
As the final step, the background color of each pixel in the color mask is computed in each of the 8 directions and the center with a median filter. If none of these medians are color, then the pixel is marked as color, so it is removed. If one of them is a color, then the saturation of the background and the color is compared, and if it is the same then the pixel is marked as text, so it is not removed.
After getting the pixels that are marks, an hysteresis is performed to detect the surrounding ones that may be color as well, but that due to the perspective, the image or the ink itself is has not enough saturation. This process has a big disadvantage: text that is "touching" color marks will be removed as well.
