The definition of quality criteria can serve to manage the expectations of digitisation processes, guide the acquisition of equipment, the selection of processing software and the selection of storage and publishing infrastructures. This section describes recommended quality criteria of research quality images. The criteria are derived from practical experience of the MBG while implementing and improving their digitisation practices, in line with established standards and recommendations.

The Global Plants Initiative guideline for herbarium specimen digitisation (JSTOR 2018a, JSTOR 2018b) establishes 600 PPI as the recommended resolution for scanning herbarium sheets. This resolution applies to the TIFF image. This image is not usually published but serves as the base from which other images versions are derived. The resolution of these images will vary according to the intended use of the derived images (such as web publishing or printing).

Table 1 describes the standard criteria for digital images of herbarium sheets. The details about the different resolutions and uses are derived from the recommendation from the Library of Congress (Library of Congress 2018), Synthesys 3 (Phillips et al. 2014), Global Plants Initiative (JSTOR 2018a, JSTOR 2018b) and the Federal Agencies Digitisation Guidelines (Federal Agencies Digital Guidelines Initiative 2016, Federal Agencies Digital Guidelines Initiative 2018). The recommendations for bit depth and colour accuracy are derived from the technical recommendation from the Library of Congress (LoC 2018, Library of Congress 2018). Finally, the recommended colour space is Adobe RGB (1998), taken from FADGI (Federal Agencies Digital Guidelines Initiative 2016, Federal Agencies Digital Guidelines Initiative 2018). The quality management criteria can be used to implement quality assessment activities within the digitisation workflow.

MBG determined that, for every specimen, a set of three images need to be produced: a high definition uncompressed archive quality master image (TIFF 450 PPI). Apart from the master image, two derivatives are produced: a high definition lossless image for derivation of other images (JPEG2000 420 PPI) and a lossy image for web publishing/online inspection (JPEG 420 PPI). The high definition uncompressed image for archiving is intended for long-time preservation, the high definition lossless image is intended to provide a working image which is easy to store, transfer and process and the low-resolution image is intended for online publishing. Any other image derivatives required can be produced as needed.

The DPI value is largely meaningless, because it relates to an arbitrary print size of the object. The Global Plants Initiative (GPI) project (Royal Botanic Gardens Kew 2019, JSTOR 2018a, JSTOR 2018b) specification of 600 PPI resolution is linked to the use of flatbed scanners, which were the main type of digitisation equipment when the recommendation was produced (between 2004 and 2009). The 600 PPI related to the highest resolution of the scanner sensor which still allowed fast digitisation and good quality reproduction, assuming that the print size of the image would be the same as the original specimen size. The 450 PPI value comes from assuming a print size where 450 camera pixel lengths would be printed in one inch of paper. With the high density sensors in the modern camera, the print size to achieve 450 ppi would be very large. Therefore, the PPI resolutions are relative to assuming different final print sizes. The camera actually has a sensor with a pixel density of about 4 MP/cm², which is many times the resolution of flatbed scanners.

In addition to the image quality requirements described above, herbarium sheet images must include a set of image elements. Image elements refer to visual elements which appear next to the herbarium sheet specimen and which are intended to help in the identification, processing and quality control. There are five elements recommended by the Global Plants Initiative (JSTOR 2018a, JSTOR 2018b): (1) Colour Chart, (2) Scale Bar, (3) Labels, (4) Barcode and (5) Institution Name. In the case of herbarium sheets, imaging all the elements may, on occasion, require more than one pass, since they may be in the form of booklets or paper sheets attached to the herbarium specimen. Fig. 1 shows the elements of the herbarium sheet specimen.

Figure 1.  

Examples of herbarium sheets and the required elements to capture. The left image corresponds to a specimen digitised during the GPI project and the one on the right is an specimen digitised during DOE!. The elements are (1) Colour Chart, (2) Scale Bar, (3) Barcode, (4) Labels and (5) Institution Name. As the images show, some elements may be combined, for instance the scale bar and institution name on the left and colour chart and scale on the right.*1

The colour chart is recommended for helping with quality control and post-processing; this can help in verifying the lighting, white balance and colour accuracy of the image. The Federal Agencies Guideline for Digitisation, the Library of Congress and Synthesys 3, recommend the use of the colour chart, referenced as colour target or colour checker (Federal Agencies Digital Guidelines Initiative 2016, JSTOR 2018a, JSTOR 2018b, Library of Congress 2018, Phillips et al. 2014). There are many types of colour charts and examples of many of them have been used by different institutions in their digitisation workflows. However, modern targets, such as those from Image Science Associates, are preferred over legacy targets (Colour Control Patches from Kodak), because they were developed for digital image creation and are made to tight tolerances (Cultural Heritage, Digital Transitions Division of Cultural Heritage 2018, Image Science Associates 2017). Object level targets of this type include a ruler and can be used for verification of colour, sharpness and scale.

Scale bar is recommended to enable the calculation of the dimensions of the specimen (Phillips et al. 2014).

Herbarium name (with or without logo) is required to quickly identify the institution holding the specimen (Phillips et al. 2014).

Labels are commonly placed next to the specimen attached to the herbarium sheet. Clear capture of labels is important for further processing and documentation of the specimens (Phillips et al. 2014).

Barcodes are identifiers used for cataloguing specimens which are also useful for linking them to digital specimens. Synthesys 3 and GPI recommend the use of barcodes as internal identifiers which are important for further documentation and linking of the physical and digital specimens (JSTOR 2018a, JSTOR 2018b, Phillips et al. 2014). There are different types of barcodes available. Line Barcodes (one dimensional), like the ones shown on Figure 1, have been used in MBG projects. Some guidelines recommend the use of two-dimensional barcodes (Zaman 2015) to prevent misidentification; however, MBG did not use these because they are not always readable without a scanner and curators also need to be able to use them for identification of physical specimens.

In the MBG collection, herbarium sheets specimens are identified by a barcode label. These labels conform either to the UPC-A or Code 128 format. In line with current practice in herbarium management (JSTOR 2018b, Phillips et al. 2014, Royal Botanic Gardens Kew 2019), the digital image files of specimens are assigned filenames based on the barcode. The check file name sub-task relies on a script that uses the ZBAR (Brown 2010) open source library to read the barcodes. The technical details of this sub-task are summarised in Table 5. The script verifies that each name conforms to MBG barcode conventions (length, structure, file type). The script works correctly for most of the images processed. However, there are cases in which two or more specimens and their barcodes are fixed to the same herbarium sheet. These cases may be flagged as incorrectly named and need to be verified manually. ZBAR can detect multiple barcodes, but a manual step is required to duplicate the image files and rename them, so that there is one image for each specimen/barcode.

At MBG, curators follow the recommendation of placing barcodes as close to the bottom right corner of the sheet as possible. If the location of the barcode is known beforehand, ZBAR can be configured to read just that part of the image, speeding up processing time considerably. However, this would only work in a collection where there cannot be two specimens on the same sheet.

Most herbarium sheets have a standard size. Consequently, herbarium sheet imaging produces images of a size which fall within a predictable range. Based on this observation, MBG established a heuristic rule correlating file size, cropping and image resolution. This has helped in establishing the accepted file size limits for the different types of image files handled and generated during image processing. The technical details of this sub-task are summarised in Table 6.

This process also includes verifying the width and height of the image which is also a dependable indicator for detecting bad crops and malformed images. The check file size and resolution sub-task relies on a script that used statistical data of past digitisation campaigns. The script collects the size of each image file and verifies that it falls within the expected range. Images outside this range are flagged as examples of bad cropping. Large files tend to be under-cropped and small files over-cropped. This automated size check can be relied on to test cropping on a large image set, but it is less sensitive than a visual cropping check. A visual cropping check can identify small cropping problems on images, but it is impractical for large datasets. Similarly, the resolution of the image in pixels is related to the file size. In a previous version, the script tested the resolution of the image in dots per inch (DPI), because it was part of the contract terms with the external contractors. However, testing the dimensions of the image in pixels is a simpler reliable test of resolution.

The MD5 algorithm is a hash function used to create a checksum for a file. It can be used to ensure that an image has not been corrupted. The hash function is created soon after the image is made and needs to be stored for later use. The check TIFF Metadata File Structure sub-task relies on a script that computes the MD5 values from the image files and verifies that it matches the value generated for the original file, shortly after digitisation (md5deep and hashdeep software packages, Table 7). Initially, validation of MD5 values was conducted on entire batch sets; however, after various digitisation campaigns in which no errors were found, the process was modified to perform spot checks on a subset of the TIFF master set, for quality control purposes. Additionally, calculation and verification of MD5 is time-consuming and applying it strategically helps speeding up the execution of the entire image processing task. In addition, the hash values are stored and passed on to the long-term archive where it is used to verify the files deposited in long-term storage. The MD5 value may be used in the future to ensure the integrity of the files in long-term storage, to verify file integrity after transmission, as well as the integrity of files which are recovered from the archive when required.

The large numbers of specimens being processed in mass digitisation operations increases the risk of duplicate imaging. Additionally, as the description of the Check File Name sub-task explains, when a herbarium sheet contains more than one specimen, the image file is duplicated manually as many times as needed, to produce at least one image for each barcoded specimen on it.

MBG created a script that searches and flags image duplicates. The script uses barcode verification and md5deep (Table 8) to search for potential duplicates in the image set being processed. This check is not difficult nor time-consuming if the MD5 values are available.

For long-term archiving, the specimen images need to conform to well-known standards, in order to improve the chances of recovery in the future. The Tagged Image File Format (TIFF) format is recommended standard for long-term storage (Ariño and Galicia 2005). Additionally, high definition images, which can be made available for wide use, are stored onsite. JPEG2000 is the format selected for storing production images on site (JP2). This format was selected because it allows storing compressed images that are lighter (i.e. smaller file size) than TIFF images, retains all the image information (lossless, no loss of information in the compression) and their quality does not degrade during generation or transmission. These images are designated production images as they are intended for publishing, sharing and generation of new derivatives. A third set of images for publishing and online inspection is generated in JPG format. This set is also part of the production set and it is used for publishing and inspection online. This set is not validated. MBG created a script that verifies the conformity of the images in the master set (TIFF) and those of the production set (JP2000) while searching and flagging image duplicates. For verifying that the images conform to the selected standards, the script uses JHOVE for TIFF images (Open Preservation Foundation 2019b) and jpylyzer (Open Preservation Foundation 2019a for JP2000 images (Table 9). The integrity of the derived JPEG files is not verified because they can always be recreated from the master (TIFF) or the high definition lossless (JP2000) images.

There is still a need to visually verify a subset of the images for quality control. This is particularly important when the digitisation process has been outsourced as this is the means for verifying that the provider meets the service level agreements. Table 10 shows the technical details of this sub-task. Additional checks to verify image quality are performed visually, the quality problems verified in this sub-task being: the focus, cropping, colour, contrast and a visual check to see if the barcode is indeed corresponding with the filing name. Visual inspection is a time-consuming process and cannot be performed for every image in large datasets. For this reason, the verification is performed in a sample, in two stages. For the first 10 batches, 2% of the images are inspected (the last of every fifty images). If the tests do not detect a problem, the set is reduced to 1% of the images in the set (the last of every 100 images). Quality problems, such as bad focus or poor lighting, tend to occur at a point in time and continue in every image until the fault is corrected. Therefore, rather than selecting images randomly, inspection at regular intervals throughout the batch is more effective to identify the point at which a problem started to occur.

Removing potential duplicates and bad crops is a process that needs to be verified visually. When some images are flagged, the processing of a batch is not stopped, i.e. the batch is not rejected if errors are detected, unless the errors are found in more than 5% of the images. Instead, the erroneous images detected are removed from the batch before storing them locally or in the long-term repository. The details of this sub-task are shown in Table 11. This is a manual task performed using the error log produced during previous verifications. It is considered a quality assurance task because it ensures that no images flagged as erroneous are saved with the rest of the batch. The rejected images are reported to the imaging team or to the contractor, for re-imaging.

The production set consists of a set of high definition JP2 images and a set of lightweight JPG images. During processing and validation, these sets are temporarily stored on a buffer server. Once the production set has been completely processed and validated, the operator executes a task to verify that all compliant image derivatives have been copied to the image repository before the buffer is cleared in preparation for processing the next batch (Table 12).

The master set contains the high definition uncompressed TIFF images which are intended for long-term storage preservation and future recovery purposes. During processing and validation, this set is temporarily stored on a buffer server. Once the master set has been completely processed and validated, the operator must verify that all compliant images have been copied to the external archive repository, before the buffer is cleared in preparation for processing the next batch (Table 13). The script for clearing the TIFF set from the buffer includes a step that verifies if the external archive provider has acknowledged the reception and archiving of the image file. If digitisation is outsourced, the script will also send a signal to the contractor to announce that the batch has been processed completely, so they can also clear their temporary storage.

The quality of the images created in-house and those outsourced are equivalent. Except for the metadata establishing which process was used for digitising, the quality criteria including image dimensions, resolution and image elements are the same. The images in Fig. 5 show examples of specimens from two herbarium sheets, digitised using in-house and outsourced digitisation, presented side by side to highlight the consistency of the results obtained. The top images (a, c) are scaled to 1/10x or 10% of their actual size. The bottom images (b, d) present 5x5 cm segments highlighted with yellow frames in "a" and "c" at a 1:1 (1x) scale. Both images were downloaded from the MBG Botanical Collections Portal (www.botanicalcollections.be) and are derivatives from the JP2000 images, encoded using the eci-RGB colour profile, 24-bit colour depth, at 420 PPI.

Figure 5.  

Example of results from in-house and outsourced digitisation. Image "a" corresponds to a specimen digitised in-house and image "b" corresponds to a specimen digitised by the contractor. Images "c" and "d" correspond to close-ups of the sections highlighted in "a" and "b", respectively, presented at 100% size (5x5 cm square).

Fig. 6 provides a further view of how the two examples from Fig. 5 compare to each other when the images are scaled. The first 1x column shows the corresponding 2x2 cm sections from Fig. 5 "b" and "d" respectively, keeping the 1:1 proportion. The remaining columns show close-up fragments of the same section scaled at 2, 4, 8 and 16 times their original size. As the images show, the magnification up to 8x is acceptable, as the borders and object features are shown clearly, with no pixelation. From 16x, the pixels are noticeable around the edges of the specimens. Magnifying physical specimens usually requires using lenses or microscopes, that only allow viewing limited sections of the specimen. However, larger high resolution displays allow viewing greater proportions of the specimen features.

Figure 6.  

Image Resolution and Scaling Comparison. HS A fragments correspond to close-ups of the images shown in Fig. 5a-c. HS B fragments correspond to close-ups of the images shown in Fig. 5b-d.

The digitisation workflow of the Royal Botanic Garden Edinburgh (RBGE) is an example of an in-house digitisation workflow. The main reasons for designing and developing this workflow in-house was the variation of funding and scale of digitisation campaigns (Haston et al. 2012). For this, a custom and modular system which could be adapted to cater for the needs of different digitisation projects was required. The RBGE designed three concurrent workflows which responded to these requirements. Fig. 7 shows the specimen, data and image workflows side by side. Notice that the image quality control activities are part of the image workflow and occur after digitisation. The workflows developed by RBGE are suitable for the characteristics of the different funding streams and the diversity of digitisation projects. The design of modular and integrated workflows to manage the processes, images and data has served this purpose. This approach is different to the process described by MBG which includes outsourcing the mass digitisation operation and external archival partners for long term storage.

Figure 7.  

Diagram of the digitisation workflows at the Royal Botanic Garden Edinburgh (from Haston 2012).

The Natural History Museum, London (NHM) has also developed a workflow for the digitisation of microscope slides, which significantly improved their digitisation efficiency, as compared to previous semi-automated methods. They also found that their automated workflow reduced the number of errors (Allan et al. 2019). Yet others have found that the modulisation of tasks in the data capture stage of processing improves efficiency and accuracy (Granzow-de la Cerda and Beach 2010).

Picturae is a digitisation provider with more than ten years of experience digitising libraries and museum collections (including cultural and natural history collections). The digitisation workflows, developed by Picturae (Fig. 8), illustrate how image quality management activities are integrated (Picturae 2017). The workflow is designed to perform image quality assessment at digitisation time, stopping and rewinding the conveyor belt if an error is detected. The means for verifying quality are built into the workflow allowing the rapid verification of every image. This workflow has been applied and tested in large digitisation projects for institutions such as Naturalis (Netherlands) and the Muséum National d’histoire Naturelle (France) and in working with MBG, as reported above.

Figure 8.  

Picturae digitisation workflow for herbarium collections. The digitisation task (3) includes image quality control checks (courtesy of Picturae*3).

The outsourced approach is effective in mass digitisation projects. There are similar examples of such projects involving other providers, such as the digitisation of the Moscow University Herbarium (Seregin 2018) or the Natural History Museum of Norway (Digitarium 2017).