Transform images

Status

What other tools call it

Meridian's classification; an independent blind rating agreed on 95.7% of 324 rows (variant, version 3).

Background

When to use it

Tesseract, a free, open-source optical character recognition (OCR) engine, does various image processing operations internally, using the Leptonica library, before the OCR itself.^[3][4] Tesseract's internal image processing will inevitably not be good enough in some cases, which can result in a significant reduction in accuracy.^[4] Tesseract works best on images with a resolution of at least 300 dots per inch (DPI), so it may be beneficial to resize images.^[4] For Tesseract 4.x, text should be dark on a light background, while Tesseract 3.05 and older handle inverted images, with light text on a dark background, without problem.^[4] The quality of Tesseract's line segmentation reduces significantly if a page is too skewed, so the page image should be rotated until its text lines are horizontal.^[4] Tesseract converts an image to black and white internally with the Otsu algorithm, and the result can be suboptimal, particularly if the page background is of uneven darkness.^[4] A big border around a single letter, digit or word can cause problems in Tesseract, such as an empty page, so the image should be cropped to the text area with a reasonable border, such as 10 pixels.^[4] Heavy ink bleeding in historical documents can be compensated for by erosion, which shrinks characters back to their normal glyph structure.^[4]

Alteryx's Image Processing tool processes its steps in order from top to bottom, and the steps can be dragged into a different order.^[5] Alteryx recommends Adaptive Gaussian thresholding in most cases for images of text, such as scanned pages of a book, and Binary thresholding for other images, such as photographs.^[5] Scaling a raster image down to fewer pixels usually results in a visible quality loss.^[6] Write down the steps in the order they run and keep the original image beside the result, so a step that harms recognition can be found and taken out. (Meridian's view) Resize to what the next step needs and no further: shrinking an image to save space can remove the detail that reading or recognising it depends on. (Meridian's view)

The method

Many of the techniques of digital image processing were developed in the 1960s at Bell Laboratories, the Jet Propulsion Laboratory, the Massachusetts Institute of Technology, the University of Maryland and a few other research facilities.^[2] In digital image processing, thresholding is the simplest method of segmenting images, and can create binary images from a greyscale image.^[7] The simplest thresholding methods replace each pixel with a black pixel if its intensity is less than a fixed threshold, or a white pixel if it is greater.^[7] Among automatic thresholding methods, Otsu's method is the most famous and widely used.^[7] Otsu's method, named after Nobuyuki Otsu, returns a single intensity threshold separating foreground from background pixels, chosen by minimising the intensity variance within each class.^[8] Tesseract 5.0.0 added two binarisation methods based on Leptonica, Adaptive Otsu and Sauvola.^[4]

Mathematical morphology was developed in 1964 by Georges Matheron and Jean Serra at the École des Mines de Paris, France.^[9] Mathematical morphology is the foundation of morphological image processing, whose basic operators are erosion, dilation, opening and closing.^[9] In RapidMiner's Dilate Image operator, dilation fills up tiny gaps in objects, so that lines get thicker and filled shapes become bigger.^[10] Image scaling can be interpreted as image resampling, and by the Nyquist sampling theorem a smaller image can be made only after a suitable 2D anti-aliasing filter prevents aliasing artefacts.^[6] RapidMiner's Align Image operator aligns an image, such as a scan of a form, to a template by detecting key points on both, matching them and calculating the transformation between them, a method called homography.^[11]

In RapidMiner, a grey scale image is a matrix in which each cell represents the intensity of a pixel, and a colour image is commonly represented by three matrices, for blue, green and red.^[12] RapidMiner's Image to Table operator turns a one-channel image into a single table of each pixel's brightness, and an image with more channels into a collection of tables, one per channel.^[13]

Compared with the alternatives

In RapidMiner's Erode Image operator, erosion is much like dilation except that it erodes the boundaries of the foreground object, diminishing the features of an image.^[14] A vector graphic can be scaled to any resolution with no loss of image quality, while scaling a raster image means generating a new image with more or fewer pixels.^[6]

Sources

1. [1]Computer vision - Wikipediasecondary source
2. [2]Digital image processing - Wikipediasecondary source
3. [3]Tesseract (software) - Wikipediasecondary source
4. [4]Improving the quality of the output | tessdocdocumentation
5. [5]Image Processingdocumentation
6. [6]Image scaling - Wikipediasecondary source
7. [7]Thresholding (image processing) - Wikipediasecondary source
8. [8]Otsu's method - Wikipediasecondary source
9. [9]Mathematical morphology - Wikipediasecondary source
10. [10]Dilate Image - Rapidminer Documentationdocumentation
11. [11]Align Image - Rapidminer Documentationdocumentation
12. [12]Table to Image - Rapidminer Documentationdocumentation
13. [13]Image to Table - Rapidminer Documentationdocumentation
14. [14]Erode Image - Rapidminer Documentationdocumentation