Regions of Interest Quantified – Solving Problems of the Analogue Methods of the Discoveries in the Judean Desert Series (Part III.2)

Our last post ended by introducing the idea of Regions of Interest (ROI). Regions of Interest is not a new idea in terms of computer science. Actually, Regions of Interest was concomitant with the definition of a digital image itself,1 but let’s not get too technical. It is better if we contextualise the matter in terms of the philological problem we face, and then raise questions how digital humanities can help us solve the problem. The problem we face is what can we (a) read in 1QSa II 11–12 given the severe damage of the parchment and (b) how can we make our method clear so that our readers understand our decisions. Before we can begin to analyse these two lines in greater detail, we need to implement some best practices in digital editorial methods.

In this post of the series, I will introduce two best practices. The first best practice is to collect all the images of any given fragment. The second best practice is to always have the ability to regress to the original images. Let’s explicate these two best practices in further detail so we can return to 1QSa II 11–12.

Best Practice 1: Collect All Images of Any Given Fragment

The first step best practice to implement relates to understanding what images we have of a given artefact, in our case 1QSa column II. But why? The reason for this was given in the last post. The artefacts have continued to decay over time, and sometimes different images of the same artefact can reveal different capta, be it material features, signs (ink traces or scribal annotations), and/or damage patterns. A capta is, therefore, understood as a taken measurement of interest, hence the feasibility of a Region of Interest. As Johanna Drucker reminds us, ”Capta is ‘taken’ actively while data is assumed to be a ‘given’ able to be recorded and observed. From this distinction, a world of differences arises.” She continues to spell out an important difference between capta and data by saying, “Humanistic inquiry acknowledges the situated, partial, and constitutive character of knowledge production, the recognition that knowledge is constructed, taken, not simply given as a natural representation of pre-existing fact.”2 If we have different images and a capta from one image is what informs our philological decisions, we need to inform our reader from what image we are making our reading.

Best Practice 2: Maintain the Integrity of the Original Image

This takes us to the second point I would like to make about best practices. We always maintain the ability to regress to the original image, because reading is ambiguous and quite complicated. It is possible that the eye can deceive, so it’s necessary to create a method that quickly enables a comparison with the original.

We need to ask how exactly are we reading an image? Here, reading means both interpreting the image with our human eyes and loading the image into a computer’s memory. Our eyes are easily tricked in certain circumstances. For example, would you say the gray hue in box A is different than than the gray hue in box B?3

Checkbox Illusion

As it turns out, the colours are the exact same. Whereas we might have an intense argument over who might be right, the computer can demonstrate that they are the same colour. The reason the computer can demonstrate that they are the same hue of gray is because it reads the image a different way than the human eye reads the image. The computer reads the image and displays the colour based on the value of the pixels at this region of the image. For the computer, it is the same pixel values.

I am not suggesting we can attain objectivity here with the computer—quiet the opposite. I am suggesting that the computer serves as a tool to help explicate the phenomenon. This is important, for it is the basis of training a neural network to read fragments as a heuristic guide for philological study of the scrolls. So what? How does this substantiate the idea that we always need the ability to regress to an original image?

This might seem irrelevant to studying Dead Sea Scrolls fragments, but consider 4Q230 frag. 1 (PAM 41.712). This was a fragment that was overlooked in the DJD editions, but later published by Eibert Tigchelaar:4

4Q230 Frag. 1 (PAM 41.712 Frag. 12)

Do you see what looks like an ink trace after the kaf on line 5? Yeah, pretty confusing, right? Let’s use the Region of Interest to get more specific here.

4Q230 Frag. 1 (PAM 41.712 Frag. 12)

In the lower left hand quadrant of the box, do you see what appears to be a shadow or an ink trace? Let’s zoom in to make this even easier:

4Q230 Frag. 1 (PAM 41.712 Frag. 12) l. 5

This appears as if this is an ink trace where perhaps part of the ink has flaked off. If a piece of ink did flake off, we could perhaps read ]חשך כי נקל[. But is this really a trace of ink? How can we tell? Thankfully, the fragment exists on other PAM images (viz., best practice number 1: find all the images of the fragment). Consider, for example, the right side of the fragment which was imaged on PAM 41.282:

If we use Python and OpenCV,5 we can do some simple computing to get the following outputs: left=original; left-center=otsu_threshold; right-center=inverted; right=canny_edge

PAM 41.282 Frag. 7

What appeared as an ink trace in PAM 41.712 seems to be either a trace of ink that was once part of the lāmed or a collection of dirt in a hair follicle in PAM 41.282. It seems as if the join has created a circumstance where our eyes see this spot as darker than it appears when not joined, which can be confirmed by the threshold image and the inverted image.

It is important to note how we have come to this insight. We came to this insight by using our first best practice, gather together all the images of one fragment.6 Second, we used some simple Python and OpenCV processing to get different ways for our eyes to read the image. Once we apply transformations to an image, we need to compare the new image with the original. This is particularly important once we get to geometric transformations, which we can use to de-warp fragments.

Before we conclude, let’s look again the Regions of Interest of PAM 41.712 Frag. 12. I would like to focus again on lines 5 and 6 of the fragment, particularly how sign 89 falls underneath the spot we have examined. Sign 89 is consistent with the geometric shape of lāmed, in which case we could tag sign 70 with both of our interpretations suggested above: either a dirt trace or a trace of the lāmed from sign 89. I will explain the idea of tagging the Region of Interest in the next post.

PAM 41.712 Frag. 12


In this post, I have introduced two ideas about best practices. The first best practice is to gather together all the images of any given fragment. These images will likely contain different information, and this information needs to be recorded (capta) and analysed in our reading of the fragment. Our reading, however, can be made better by using the computer as a tool. As Steve Jobs said, “The computer is a bicycle for the human mind.” By this, Jobs meant that we can become more efficient with the computer. We can manipulate images in ways to bring out features with which the human eye struggles, but we need to make this clear. We always need to have the ability to compare the manipulated image with the original. Now that I have introduced these ideas, we can return to 1QSa II 11–12 in the next post and explore some of the proposed readings, and perhaps even suggest a new reading!

  1. As far as I am aware, I was the first to bring the idea of Regions of Interest into digital analysis of the Dead Sea Scrolls. I presented a paper on the benefit of using Regions of Interest in my study of the Isaiah Scrolls. I gave this paper at a conference in Copenhagen in early 2014. I made a prototype digital edition of the Isaiah Scrolls, and later shared this prototype with Reinhard Kratz in Göttingen, on 14 April 2014. Reinhard Kratz forwarded my prototype to Shani Tzoref, Ingo Kottsieper, and Annette Steudel on 23 April 2014. In later discussions with Shani Tzoref, she informed me that she used my prototype edition to write an application for a digital edition of the Damascus Document and an application for Scripta Qumranica Electronica. This is the reason why I was hired in Göttingen to work on the project.
  2. Johanna Drucker, “Humanities Approaches to Graphical Display,” DHQ 5.1 (2011); online:
  3. Image used with Permission;
  4. Eibert J.C. Tigchelaar, “”These Are the Names of the Spirits of… “: A Preliminary Edition of 4QCatalogue of Spirits (4Q230) and New Manuscript Evidence for the Two Spirits Treatise (4Q257 and 1Q29a).” Revue de Qumrân 21.4 (2004): 529–47.
  5. Python is a very popular programming language. OpenCV is a package within Python that can be used for Computer Vision computational needs.
  6. There are more PAMs to include here, but I have not included in this example. I did not include them in this example because I wanted to show how the best practices are related in practice. We will look at all the examples of 1QSa II 11–12 in the next post.

Leave a Reply

Your email address will not be published.