Pith. sign in

REVIEW 3 major objections 6 minor 23 references

Intelligent Recognition of Time Stamp Characters in Solar Scanned Images from Film

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read A convolutional neural network reads hour and minute characters on scanned solar film frames, completing date/time metadata for more than 7 million images from a 1963–2003 archive at about 98% accuracy.

desk verdict A useful applied OCR pipeline that actually produced timestamps for 7.7 million NSO film frames; the method is standard, the accuracy claim is plausible but under-verified, and the paper needs a few clarifications before I'd trust the 97.9% figure. read the letter →

arxiv 1909.00316 v1 pith:7QR2PDGG submitted 2019-09-01 astro-ph.SR astro-ph.IM

classification astro-ph.SRastro-ph.IM
keywords solarimagestimestamprecognitionconvolutionalneuralnetworkcharactersegmentationfilmdigitizationhistoricalobservationsopticaldeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a Convolutional Neural Network (CNN) can read the date and time characters stamped on digitized solar film images quickly and accurately enough to make a huge historical archive usable. The authors build a two-part pipeline: a segmentation stage that locates the stamp, isolates single characters, and resizes them to $28\times28$; and a CNN classifier that identifies each character. They report applying it to more than seven million frames recorded between 1963 and 2003, with the hour and minute read automatically at a per-image accuracy of 97.9% on a randomly selected, manually checked test set of 10,000 images. Dates are fixed by manually entering one date per observing day and then using continuity rules, so the automatic part is the character recognition itself. If the accuracy holds across the archive, the payoff is a decades-long solar film record converted into machine-searchable, time-tagged data for flare and activity studies.

What carries the argument

The load-bearing mechanism is the CNN itself: two convolutional layers with $5\times5$ kernels (6 and 10 kernels, respectively), two max-pooling layers, one fully connected layer, and a softmax classifier, operating on $28\times28$ grayscale character crops with ReLU activations. It is embedded in a segmentation pipeline tailored to the stamps: the solar disk is removed, the stamp region is located and rotated, noise is suppressed with a top-hat operation, the region is binarized with the Sauvola algorithm, connected components with areas in a fixed range are kept, character regions are extracted by a stroke-width transform, and each character is resized to $28\times28$. The pipeline defaults to white characters and retries once with inverted colors when no regions are found, handling both black-on-white and white-on-black stamps. A final date-check step uses manual entry for the first date of an observing run and continuity rules to assign years, months, and days.

What would settle it

Run the trained pipeline on a held-out set of 10,000 frames from years or film types that contributed no training characters and compare the per-image accuracy with the reported 97.9%; a large drop would falsify generalization. Alternatively, inspect the 10,000-image test set for overlap with the 100,000 training crops—any duplicated character image would invalidate the reported accuracy estimate.

Watch

Extended reading notes

Core claim

The central discovery claimed in the paper is that a comparatively small CNN, trained on 100,000 manually labeled single-character crops (10,000 per digit), recognizes the hour and minute characters in these noisy, unevenly lit film stamps with per-character accuracy above 97% and an average around 99.5%. On 10,000 full images drawn at random from the archive, 9,788 were read with no character errors, 202 with one error, and 10 with two errors, which the authors report as 97.9% image-level accuracy at an average of 0.09 seconds per picture. The date fields are not read automatically; the pipeline records the date of the first picture of an observing day by hand and propagates it forward using the sequential logic of 24-hour time. Summing the per-year counts, the paper reports date/time information for a total of 7,760,911 frames, leaving unreadable or stamp-less images aside.

Load-bearing premise

The result stands on the assumption that the 10,000 manually checked test images are independent of the 100,000 training characters and representative of all 7.76 million frames, including damaged and unevenly lit stamps; if the test set is not independent or under-represents hard cases, the 97.9% accuracy will not generalize.

Editorial extensions

If this is right

  • The 1963–2003 film record becomes searchable by date and time, so researchers can automatically pull frames around known flares, Moreton waves, and other transient events.
  • The reported throughput of about 0.09 seconds per image means re-running the recognition is cheap enough for reprocessing if the archive is re-digitized at higher resolution.
  • Because the date is fixed by continuity, a single manual date entry per observing day removes the need to recognize small, blurry year/month/day characters, lowering the error budget to the hour/minute digits.
  • The confusion matrix indicates most remaining errors concentrate in characters '2', '3', and '6' under uneven lighting, so future improvements need only target those failure modes rather than the whole classifier.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported 97.9% is an upper bound unless the 10,000 test images were excluded from the 100,000 training crops; checking this overlap is a two-step audit that would settle the main uncertainty.
  • The roughly 10% of frames without readable stamps are likely not random: stamps may be unreadable precisely during bright flares or poor observing conditions, so studies of those events could be biased by the missing frames.
  • A testable extension is to add illumination-degraded character crops to the training pool and re-measure the confusion matrix; the paper itself says this is the planned fix, and the expected effect is to reduce the '2'/'3'/'6' errors.
  • The same segmentation-plus-CNN recipe is portable to other historical film archives, but the portability claim would need validation on a separate test set of at least a few thousand frames from each new archive.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents a CNN-based pipeline for automatically extracting date/time stamps from digitized 35-mm solar film images of the National Solar Observatory (NSO). The pipeline segments stamp characters via binarization, connected-component filtering, and stroke-width transform, then classifies hour/minute characters with a small CNN trained on 100,000 manually labeled single-character crops. The authors report 97.9% image-level accuracy on 10,000 randomly selected original images, compare favorably against Tesseract, and state that date/time information was extracted for 7,760,911 images from 1963 to 2003, with roughly 10% of archive frames left unprocessed because stamps were unreadable or absent. The central claim is that the method yields automatic time-stamp metadata at approximately 98% accuracy across this large historical dataset.

Significance. If the claimed accuracy holds, the work has clear practical value: it would rescue the time-stamp metadata for a large, scientifically important historical solar archive, enabling time-resolved studies of flares and other phenomena. The manuscript also provides a useful comparison against Tesseract and a concrete description of the segmentation and recognition pipeline, and the reported validation is based on manually checked samples rather than being circular. The main limitations are not in the novelty of the CNN architecture but in the precision with which the reported accuracies are defined and measured: test-set independence from the training crops is not demonstrated, the treatment of unreadable/no-stamp frames in the test set is unspecified, and there is an internal inconsistency between two accuracy tables. These issues are fixable and do not invalidate the approach, but they must be resolved before the headline accuracy claim can be accepted.

major comments (3)
  1. [Sections 3.2 and 4] The manuscript trains on 100,000 single-character images cut from original NSO frames (Section 3.2) and then tests on 10,000 randomly selected original images (Section 4), but it never states that the 10,000 test images were excluded from the source frames used to build the training pool. If any test frame contributed characters to the training set, the reported 97.9% image-level accuracy is inflated by direct overlap. Please state explicitly whether the test images were disjoint from the training frames, and if so, describe how this was ensured.
  2. [Section 4 and Table 5] The abstract claims 'more than 7 million images with the accuracy of 98%,' but the paper states that about 10% of the archive (e.g., 1971, 1986, 1990) was not processed because stamps were beyond human recognition or absent. It is unclear whether the 10,000-image test set includes such frames. If it excludes them, then the 97.9% figure applies only to the conditionally readable subset, and the end-to-end success rate on all digitized frames is approximately 0.9 × 0.979 ≈ 88%, not 98%. Please clarify whether the test set includes unreadable/no-stamp frames, and either qualify the headline accuracy as conditional on successfully extracted stamps or report an end-to-end rate over the full archive.
  3. [Tables 3 and 4] Tables 3 and 4 are internally inconsistent: Table 3 reports 202 images with one character error and 10 images with two errors, implying 222 misread characters, whereas the column error counts in Table 4 sum to 218. This discrepancy of four characters affects the derived per-character recognition rates and needs to be reconciled, either by correcting the counts or by explaining how images with multiple errors are tabulated in Table 4.
minor comments (6)
  1. [Title] The title contains a typo: 'S canned' should be 'Scanned'.
  2. [Section 2] 'sigmod' should be 'sigmoid', and 'ReLu' is conventionally written 'ReLU'.
  3. [Section 3.1, Step 5] The phrase 'Reserve connectivity domain of which area is in (500, 1000)' is unclear; it should read 'Preserve connected components whose area is in the range (500, 1000)'.
  4. [Table 4] The heading 'Number of erros' contains a typo and should read 'Number of errors'.
  5. [Reference list] The reference 'Sauvola & Pietikinen 2000' misspells the second author's name; it should be 'Pietikäinen'.
  6. [Section 3.3] The date-check GUI is described qualitatively; it would be helpful to state how many dates required manual correction and how such corrections propagated through the archive, since this affects the reliability of the final date metadata beyond the character-recognition accuracy.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: accuracy is an empirical measurement against manually labeled data, not a derived quantity.

full rationale

The paper's accuracy claims (Tables 1, 3, 4) are reported measurements on test images checked by humans, not quantities derived from the model's training objective. The CNN is trained on labeled character images and then evaluated on 10,000 randomly selected original images; the 97.9% figure is a measured outcome, not an identity. The citation to Zheng et al. (2016) is used only to motivate the choice of CNN and does not supply the accuracy numbers. The comparison against TESSERACT provides an external benchmark. A possible overlap between the 100,000 training crops and the 10,000 test images is not stated and would be a data-leakage concern, but the text provides no evidence of such overlap and this is a methodological risk, not circularity. No equation in the paper defines a target quantity in terms of an input in a way that would make the 'prediction' true by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method depends on hand-chosen segmentation thresholds and CNN architecture, plus the assumption that archive images can be cleanly separated into readable characters. No new physical entities are introduced, and the date resolution is partly manual.

free parameters (5)
  • CNN architecture = Con1: 6 5x5 kernels; Con2: 10 5x5 kernels; max-pool 2x2; softmax output
    Hand-selected architecture with no ablation or search reported; recognition accuracy depends on these choices.
  • Character connected-component area range = 500 to 1000 pixels
    Step 5 keeps only components in this area range to isolate characters; the threshold is ad hoc for this film format and not justified.
  • Character resize size = 28 x 28 pixels
    Standard CNN input size adopted without justification; affects segmentation alignment and recognition.
  • Sauvola binarization parameters = Unspecified window size and k
    Step 4 invokes Sauvola's algorithm but omits the window and k parameters needed to reproduce the segmentation.
  • Top-hat structuring element = Unspecified
    Noise removal in Step 3 is not specified, so the segmentation cannot be re-implemented exactly.
assumptions (5)
  • domain assumption Time stamps appear only on the left or right side of each image and are either black-on-white or white-on-black.
    Used in Steps 1-7 to localize and binarize; if other layouts exist, the pipeline will miss them.
  • domain assumption Characters are cleanly separated by connectivity and stroke width after binarization, with area in the 500 to 1000 pixel range.
    Steps 5-6 assume isolation into connected components; the paper itself notes that illumination can break characters, as seen in Figure 10.
  • domain assumption The 100,000 training characters cut from original images are representative of, and disjoint from, the test images; the manual labels are correct.
    No leakage prevention or independent label verification is reported; this assumption directly supports the reported accuracy.
  • ad hoc to paper Observation dates are continuous within folders and can be propagated from a manually entered first date using 24-hour time rules, with manual checks for gaps.
    Section 3.3 date check: for example, a '2359' frame followed by '000' implies the next day; gaps require manual intervention.
  • standard math The Sauvola binarization, stroke width transform, and CNN backpropagation algorithms work as described in the cited literature.
    Segmentation and training rely on external algorithms that the paper does not re-derive.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Intelligent Recognition of Time Stamp Characters in Solar Scanned Images from Film." pith.science (2026). https://pith.science/paper/7QR2PDGG

@misc{pith2026190900316,
  author       = {Pith},
  title        = {Pith review of: Intelligent Recognition of Time Stamp Characters in Solar Scanned Images from Film},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7QR2PDGG}},
  note         = {Machine review of arXiv:1909.00316}
}
read the original abstract

Prior to the availability of digital cameras, the solar observational images are typically recorded on films, and the information such as date and time were stamped in the same frames on film. It is significant to extract the time stamp information on the film so that the researchers can efficiently use the image data. This paper introduces an intelligent method for extracting time stamp information, namely, the Convolutional Neural Network (CNN), which is an algorithm in deep learning of multilayer neural network structures and can identify time stamp character in the scanned solar images. We carry out the time stamp decoding for the digitized data from the National Solar Observatory from 1963 to 2003. The experimental results show that the method is accurate and quick for this application. We finish the time stamp information extraction for more than 7 million images with the accuracy of 98\%.

Figures

Figures reproduced from arXiv: 1909.00316 by the authors.

Figure 1
Figure 1. Full-disk chromosphere image obtained in Hα line by NSO. The time stamp is on the left side of the image, and the full-disk chromosphere image in the middle [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Convolutional neural network structure, with input layer, convolution layer, pooling layer, fully connected layer, and output layer. Multiple convolutional layers, pooling layers, and fully connected layers are possible in the CNN. The convolution layer detects the characteristics of the input layer to the maximum extent by ran￾domly generating sufficient convolution kernels. A large number of feature maps are gener… view at source ↗
Figure 3
Figure 3. Black character image [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: White character image. The left part of the flow chart introduces image segmentation, and the right part introduces char￾acter recognition. The input image is processed by white characters by default. If no character areas can be extracted, the process returns to the b…
Figure 5
Figure 5. Figure 5: Algorithm flow chart. that there are no characters in the current picture. Because there are only two forms of time stamp and few a part of the images that do not contain time stamps, the time stamp characters cannot be extracted from these images during the above proc…
Figure 6
Figure 6. Figure 6: Extract characters from [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Convolution neural network structure of character recognition, included a input layer, two convolution layers, two pooling layers, and a fully connected layer. The training steps of the CNN in this paper are divided into the following three steps. Step 1. Add labels to…
Figure 8
Figure 8. Figure 8: It is hard for OCR to handle these situations. So it can be concluded from the comparative experiments that CNN has better robustness, stronger antijamming, and lower time consumption than OCR [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Graphic interface of date check [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: When they are affected by illumination, they are easily destroyed by the local binary algorithm leading to structural breaks. The character fragments are considered to be noise in the next step of the algorithm because of their small area, which will affect the recogn…
Figure 11
Figure 11. Figure 11: Original image (a, c) and segmentation result (b, d), respectively. As shown in figure (b, d), the recognition result is correct, even though the fourth character is partially split [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Number of pictures per year. 5. CONCLUSION In this paper, we describe an intelligent algorithm to extract the time stamp from traditional films based on CNN. The experimental results show that the method has a good result and meets the speed and quality requirements f…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 8 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

  3. [3]

    PelrkSQKs1TGb]

    thebibliography [1] 20pt to REFERENCES 6pt =0pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key on reference command Each re...

  4. [4]

    1986, Acta Astronomica Sinica, 27, 173

    Ai , G.-X., & Hu , Y.-F. 1986, Acta Astronomica Sinica, 27, 173

  5. [5]

    2010, in 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2963--2970, 10.1109/CVPR.2010.5540041

    Epshtein , B., Ofek , E., & Wexler , Y. 2010, in 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2963--2970, 10.1109/CVPR.2010.5540041

  6. [6]

    J., Bulatov, Y., Ibarz, J., Arnoud, S., & Shet, V

    Goodfellow, I. J., Bulatov, Y., Ibarz, J., Arnoud, S., & Shet, V. 2013, in In ICLR 2013. 1312.6082

  7. [7]

    2018, Pattern Recognition, 77, 354 , 10.1016/j.patcog.2017.10.013

    Gu, J., Wang, Z., Kuen, J., et al. 2018, Pattern Recognition, 77, 354 , 10.1016/j.patcog.2017.10.013

  8. [8]

    2012, IEEE Signal Processing Magazine, 29, 82, 10.1109/MSP.2012.2205597

    Hinton , G., Deng , L., Yu , D., et al. 2012, IEEE Signal Processing Magazine, 29, 82, 10.1109/MSP.2012.2205597

Show all 23 references
  1. [9]

    2014, CoRR

    Kim, Y. 2014, CoRR. 1408.5882

  2. [10]

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2017, Commun. ACM, 60, 84, 10.1145/3065386

  3. [11]

    L., Ah Chung Tsoi , & Back , A

    Lawrence , S., Giles , C. L., Ah Chung Tsoi , & Back , A. D. 1997, IEEE Transactions on Neural Networks, 8, 98, 10.1109/72.554195

  4. [12]

    2015, Nature, 521, 436, 10.1038/nature14539

    LeCun, Y., Bengio, Y., & Hinton, G. 2015, Nature, 521, 436, 10.1038/nature14539

  5. [13]

    2012, in Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), 681--684

    Li , Y., & Lu , H. 2012, in Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), 681--684

  6. [14]

    2013, The Astrophysical Journal, 773, 166, 10.1088/0004-637x/773/2/166

    Liu, R., Liu, C., Xu, Y., et al. 2013, The Astrophysical Journal, 773, 166, 10.1088/0004-637x/773/2/166

  7. [15]

    M., & Mao, J

    Mohiuddin, K. M., & Mao, J. 1999, Optical Character Recognition (American Cancer Society), 226--236, 10.1002/047134608X.W5502

  8. [16]

    Nasien , D., Haron , H., & Yuhaniz , S. S. 2010, in 2010 Second International Conference on Computer Engineering and Applications, Vol. 1, 249--252, 10.1109/ICCEA.2010.56

  9. [17]

    2000, Pattern Recognition, 33, 225 , 10.1016/S0031-3203(99)00055-2

    Sauvola, J., & Pietikäinen, M. 2000, Pattern Recognition, 33, 225 , 10.1016/S0031-3203(99)00055-2

  10. [18]

    2015, Neural Networks, 61, 85 , 10.1016/j.neunet.2014.09.003

    Schmidhuber, J. 2015, Neural Networks, 61, 85 , 10.1016/j.neunet.2014.09.003

  11. [19]

    2007, in Proceedings of the Ninth International Conference on Document Analysis and Recognition - Volume 02, ICDAR 2007 (Washington, DC, USA: IEEE Computer Society), 629--633

    Smith, R. 2007, in Proceedings of the Ninth International Conference on Document Analysis and Recognition - Volume 02, ICDAR 2007 (Washington, DC, USA: IEEE Computer Society), 629--633. http://dl.acm.org/citation.cfm?id=1304596.1304846

  12. [20]

    2014, in 2014 IEEE Conference on Computer Vision and Pattern Recognition, 1891--1898, 10.1109/CVPR.2014.244

    Sun , Y., Wang , X., & Tang , X. 2014, in 2014 IEEE Conference on Computer Vision and Pattern Recognition, 1891--1898, 10.1109/CVPR.2014.244

  13. [21]

    2012, The Astrophysical Journal, 760, 101, 10.1088/0004-637x/760/2/101

    Wang, H., & Liu, C. 2012, The Astrophysical Journal, 760, 101, 10.1088/0004-637x/760/2/101

  14. [22]

    H., Frank, E., Hall, M

    Witten, I. H., Frank, E., Hall, M. A., & Pal, C. J. 2017, Data Mining: Practical Machine Learning Tools and Techniques, 4th edn. (Morgan Kaufmann), 10.1016/C2015-0-02071-8

  15. [23]

    2016, New Astronomy, 45, 54 , 10.1016/j.newast.2015.11.001

    Zheng, S., Zeng, X., Lin, G., et al. 2016, New Astronomy, 45, 54 , 10.1016/j.newast.2015.11.001

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.