In duncantl/Rtesseract: Interface to the tesseract OCR system

Consider the image

When we perform OCR on this, tesseract treats the highlighted text in "yellow" as a dark rectangular block and ignores it. Let's see what we get from tesseract() with this image:

t0 = tesseract("../images/SMITHBURN_1952_p3.png")
w0 = GetText(t0)

Some of the words in the highlighted text are lot, 65 and arboreal. Are these in the words returned by tesseract:

c("lot", "65", "arboreal") %in% w0

The word lot is. This is found in the fourth line, e.g. "a lot of 86 Aedes"

This suggests counting the number of occurences of each word:

tt0 = table(w0)
tt0[c("lot", "65", "arboreal")]

So we did get one instance of lot, but not the one under the highlighted text.

Of course, there is text under this highlight. We need to pre-process the image before we pass it to tesseract to have the possibility of recovering this text. We do this using some of the functions that interface to the leptonica library. In this case, it turns out we need only convert the image from color to gray-scale or 8 bits per pixel. The tesseract will be able to recover the text. We covert the image to gray-scale with the pixConvertTo8() function. We read the image and convert it with

pix = pixRead("../images/SMITHBURN_1952_p3.png")
pix1 = pixConvertTo8(pix)

We can then pass the modified image to the tesseract() function:

t1 = tesseract(pix1)

Then we call, e.g., GetText() as before, e.g.,

w1 = GetText(t1)

We then check whether the words we see under the highlighted text are in the words returned by tesseract for this modified image:

c("lot", "65", "arboreal") %in% w1

Indeed all three are there. And if we cound the number of occurrences of each

table(w1)[c("lot", "65", "arboreal")]

we see that we get a count of 1 for "65" and "arboreal" and 2 for "lot". So we got both "lot" words.

Of course, there is no guarantee that tesseract will correctly identify the words in general. We know that it did this correctly for our image. It could of course recognize arboreal as something slightly different and so not appear in the table of words.

But the lesson here is that preprocessing the image can help, and sometimes it is quite simple.

Notes

Preprocessing: why is it important, how you diagnose, what do/can you do?

1. Process of showing that the results of interest is not in the output - process of debugging results.

2. Issue: How do you know what preprocessing is needed? Point to the wiki. You need to know the image. Grey scale, binary/thresholding, subsetting/cropping, deskew/rotation, set pixels to certain values.

3. Can read in image in Leptonica, process, and then hand directly to Tesseract.

Removing at the correct level separates text from background - Tesseract is filling in gaps, creating box.

duncantl/Rtesseract documentation built on March 25, 2022, 5:50 a.m.