Pith. sign in

REVIEW 3 major objections 6 minor 23 references

Construction of efficient detectors for character information recognition

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

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A modified Census transform gives the most efficient railway-digit detectors.

desk verdict A modest, honest engineering comparison of Haar/LBP/Census features for railway carriage number detection; the headline claim outruns a single 274-image test set. read the letter →

arxiv 1908.04634 v1 pith:G2YIFXHK submitted 2019-08-13 cs.CV

A.A. Telnykh , I.V. Nuidel , Yu.R. Samorodova This is my paper · ORCID
classification cs.CV
keywords railwaycarriagenumberdetectionmodifiedCensustransformnon-localbinarypatternAdaBoostcascadeHaarfeatureslocalpatternscharacterrecognitiondigit
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

The paper claims that for locating and reading railway-carriage numbers in video, the most efficient detector features come from a modified Census transform: a 9-bit non-local binary pattern coded from a rectangular image fragment. Combined with AdaBoost learning and a cascade of weak classifiers, these features yield detectors that need fewer features and lower false-rejection rates than detectors built on Haar features or local binary patterns. The authors build 11 detector types (car-number plus digits 0 through 9), train and test on 1139 manually labeled frames, and report that the best scanning overlap parameter lies in the interval [0.75, 0.8]. If correct, the result gives a practical, low-cost choice of feature family for character recognition in video surveillance and a template for comparing feature families in other object-detection tasks.

What carries the argument

The central object is the modified Census transform extended to a rectangle of arbitrary size, which the paper calls a non-local binary pattern: a 9-bit code $C$ whose bits are set by comparing the mean brightness of each of nine equal sub-regions of a rectangular aperture with the mean brightness of the whole aperture. This turns the image into a high-dimensional feature space of codes $0 \le C \le 511$. The argument is carried by pairing that coding with the AdaBoost cascade: weak classifiers are built from per-code likelihood estimates, strong classifiers are weighted sums of weak classifiers, and detectors are serial cascades of strong classifiers; the overlap parameter controls how densely the scanning window samples the training and search images. The mechanism explains the efficiency gain as richer local structure captured in one 9-bit pattern, so the boosted cascade reaches its error target with fewer weak classifiers.

What would settle it

To test the ranking, run the same three feature families under identical cascade training on several independent datasets of car numbers, license plates, and text captured in different lighting, weather, and camera positions, and report feature counts and false-rejection rates with cross-validation. If Haar features or LBP match or beat the modified Census transform in both feature count and FRR at overlap 0.75-0.8 on any such dataset, the paper's central conclusion is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a ranking of feature families under one fixed boosted-cascade protocol: the modified Census transform is the most effective feature for forming detectors of railway carriage numbers and digits. Each image fragment is encoded as a 9-bit integer $C \in \{0,\dots,511\}$ by comparing the average brightness of each of nine sub-regions of a rectangle with the rectangle's overall average; this yields a non-local binary pattern. A weak classifier uses the estimated code histograms for signal versus background under a maximum-likelihood rule, and AdaBoost combines them into a strong classifier. Across 198 trained detectors at a false-acceptance rate near $5\times10^{-5}$, Census-based detectors consistently need the fewest features and give the lowest false-rejection rate, while Haar features require the most features and reject the most true objects; the authors conclude that the best overlap of the scanning window with the labeled object lies in $[0.75, 0.8]$.

Load-bearing premise

The load-bearing premise is that the single set of 1139 manually labeled railway-car frames, split once into training and test parts, fairly represents the full range of objects and imaging conditions, so the observed feature ranking is a property of the features rather than of this one dataset.

Editorial extensions

If this is right

  • Census-based detectors reach a false-acceptance rate around $5\times10^{-5}$ with the smallest number of weak classifiers among the three feature families.
  • For both whole-number detection and digit recognition, Haar features are consistently the weakest, needing the most features and producing the highest false-rejection rate.
  • The overlap parameter around $0.75$-$0.8$ is recommended when designing scanning detectors for objects of a given type.
  • Digit '1' is an exception where the standard $12\times24$ aperture is too large; the paper implies that a smaller aperture is needed for narrow digits.
  • The general recipe is claimed to be universal, applying to faces, vehicles, letters, and other fixed-type objects, not just railway numbers.

Reading between the lines

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

  • Beyond the paper: the same feature-ranking protocol could be run on independent datasets of license plates or text images captured in different conditions; if the Census advantage persists there, the 'universal' claim would be much stronger.
  • The paper's conclusions rest on a single annotated dataset split once without cross-validation or confidence intervals, so a reader should treat the feature ranking as provisional until it is reproduced on additional data.
  • The 9-bit code space of 512 patterns is small enough that the learned distribution of Census codes for signal versus background could be visualized, offering a principled way to see which patterns the cascade actually relies on.
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 proposes a cascade detection scheme, based on the Viola-Jones framework with AdaBoost, for detecting railway carriage numbers and recognizing digits. The main technical ingredient is a modified Census transform that encodes a rectangular image fragment as a 9-bit non-local binary pattern, compared empirically against Haar features and local binary patterns (LBP). The authors construct 11 detector types (one for carriage numbers, ten for digits), evaluate false rejection rates (FRR) and feature counts on a single dataset of 1139 frames split 3:1 into training and test sets, and conclude that the modified Census transform yields the fewest features and the lowest FRR, with the best overlap parameter in [0.75, 0.8], and that the approach is universal for searching objects of a given type.

Significance. If the empirical finding holds, the paper provides a useful data point for practitioners building cascade detectors for character information: a non-local Census-based binary pattern can outperform Haar and LBP features in feature efficiency and FRR on a specific railway carriage dataset. The paper is honest in reporting per-digit results, making public the dataset (ref. [36]), and acknowledging the failure on digit '1'. However, the significance is limited by the single-domain, single-dataset evaluation, the absence of statistical validation, and the lack of comparison with modern deep-learning baselines, which the authors explicitly defer to future work. The strength of the claim 'most effective' and 'universal' exceeds what the evidence supports.

major comments (3)
  1. [§4.1, Fig. 5, Table 1] The central comparative claim that modified Census (CS) features are the most efficient (fewest weak classifiers, lowest FRR) is based on a single random split of 1139 images into 865 training and 274 test images, with no confidence intervals, cross-validation, or significance tests. With only 274 test frames, the FRR differences between CS and LBP in Fig. 5 could plausibly be sampling noise, and the per-digit FRRs in Table 1 (e.g., 31.13% for '1' vs. 0.66% for '9') have wide binomial uncertainty. The paper should report error bars or at least a significance test for the CS-vs-LBP comparison before claiming that CS is 'most effective'.
  2. [§5, Conclusions] The conclusion that the modified Census transform is 'most effective for forming the features of objects to be recognized' generalizes beyond the evidence. The experiments cover only railway carriage numbers and digits from one dataset, and the authors themselves note that the result does not hold for digit '1'. The 'universal approach' claim in the abstract and introduction is not supported by experiments on faces, vehicles, or other object classes. This overreach should be removed or explicitly scoped to the tested domain.
  3. [§4.2, Table 1] The digit '1' detector has an FRR of 31.13%, which is far above all other digits and would be unacceptable in practice. The paper acknowledges this and suggests a smaller aperture (8x24) as future work, but this exception is load-bearing for the claim that CS features are best for digit recognition: if the most difficult digit is excluded, the conclusion covers only 'the majority of the digits'. The paper should either present results with the corrected aperture for digit '1' or clearly state that the general claim does not apply to all digits.
minor comments (6)
  1. [Abstract and §5] The phrase 'universal approach' is used repeatedly, but the experiments are limited to a single object domain. This is a presentation issue: replacing 'universal' with 'general-purpose for a given object class' would align the wording with the evidence.
  2. [§1, References] Several references are incomplete or inconsistently formatted, e.g., ref. [2] gives a journal title without volume/pages, and ref. [8] is listed with an ISBN but no publisher. The authors should verify the reference list against the journal style.
  3. [§4.1, Eq. (1)-(3)] Equations (1) and (2) are poorly typeset and hard to read due to missing summation limits and garbled indices. Please rewrite these formulas clearly, as they define the core non-local Census code.
  4. [§4.1, Fig. 5] Fig. 5 is described in the text as showing FRR versus overlap, but the figure contains multiple curves without error bars or point markers. Adding markers and error bars would make the comparison interpretable.
  5. [§4.2, Fig. 7] The text says 'Fig. 7' shows the FRR for digit detectors, but the figure is not referenced in the main body before the conclusions. Please add an explicit pointer and a short discussion of the per-digit FRR curves.
  6. [§4.1, Terminology] The paper interchangeably uses 'neural network' for the cascade of AdaBoost strong classifiers (e.g., in §4.1 and Fig. 4). Since the method is not a neural network, this terminology is misleading and should be corrected to 'cascade detector'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claims are empirical comparisons on a held-out test split, with external feature definitions and an external dataset.

full rationale

The paper's central claim—that the modified Census transform yields the fewest features and lowest false rejection rate—is an empirical result obtained by training cascade detectors on 865 frames and evaluating them on a disjoint 274-frame test set (Section 4.1). The features compared (Haar, LBP, modified Census) are defined by external references [1,2,5,6,7], not derived from the target conclusion. The dataset [36] is a collected corpus of railway-car images; its train/test split provides genuinely independent evidence about detector performance. Weak classifiers are built from probability density estimates on the training set only, and the final FRR/FAR values are measured on the test set, so no fitted parameter is renamed as a prediction. The only self-citations are references [28–31], which merely place the work in the authors' research line and carry no load-bearing mathematical or empirical weight. The post-hoc selection of the overlap range [0.75, 0.8] from the same test data is a statistical limitation (lack of error bars and multiple-comparison control), not a circularity: the conclusion is not equivalent to an input by construction. The generalization from railway digits to arbitrary objects is an overstatement given the single domain, but overgeneralization is a correctness/robustness concern, not a circular-derivation concern. The derivation chain is self-contained in the sense that every stated result is supported by a reproducible experiment on held-out data rather than by definition or by the authors' own unverified prior theorems.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing assumptions are standard machine learning background plus the representativeness of a single manually annotated dataset. The free parameters are training choices and a decision threshold learned from data.

free parameters (4)
  • Overlapping threshold = [0.75, 0.8] chosen as best from experiments
    The paper selects the overlap between labeled object and detector scanning window as a tunable parameter and concludes the best range from the experimental curves (Section 4.1, Section 5).
  • Detector aperture size for digits = 12x24 pixels (8x24 suggested for digit 1)
    The aperture is set manually, and a smaller aperture is proposed for digit 1 after observing 31% FRR (Section 4.2).
  • Strong classifier decision threshold Theta = learned per detector
    The threshold in Eq. (6) is determined during AdaBoost training and is central to the classifier output.
  • Number of weak classifiers = varies by feature type and overlap
    The paper reports the number of features as a complexity metric; the counts are a result of training, not an input, but comparisons depend on them.
assumptions (5)
  • standard math AdaBoost yields a strong classifier from weak classifiers (refs [3,4]).
    Used to build strong classifiers in Section 3.2; no proof is given in the paper.
  • standard math Viola-Jones cascade detection reduces false positives when weak classifiers are arranged in stages (refs [1,2]).
    The detector design in Section 3.3 relies on the cascade technique.
  • domain assumption Probability densities of integer codes can be estimated by histograms from a training sample.
    Section 3.1 uses histogram estimates in the maximum-likelihood decision rule; small samples for digit 1 make this unreliable.
  • domain assumption The manually labeled railway carriage dataset is representative of real video conditions.
    All conclusions in Section 4 depend on this dataset alone; no external validation is provided.
  • domain assumption The modified Census transform produces discriminative features for characters.
    The method assumes average brightness comparisons over nine sub-regions encode useful structure; no proof of discriminative power is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Construction of efficient detectors for character information recognition." pith.science (2026). https://pith.science/paper/G2YIFXHK

@misc{pith2026190804634,
  author       = {Pith},
  title        = {Pith review of: Construction of efficient detectors for character information recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2YIFXHK}},
  note         = {Machine review of arXiv:1908.04634}
}
read the original abstract

We have developed and tested in numerical experiments a universal approach to searching objects of a given type in captured video images (for example, people's faces, vehicles, special characters, numbers and letters, etc.). The novelty and versatility of this approach consists in a unique combination of the well-known methods ranging from creating detectors to making decisions independent of the type of recognition objects. The efficiencies of various types of basic features used for image coding, including the Haar features, the LBP features, and the modified Census transformation are compared. A combination of the modified methods is used for constructing 11 types of detectors of the number of railway carriages and for recognizing digits from zero to nine. The efficiency of the constructed detectors is studied.

Figures

Figures reproduced from arXiv: 1908.04634 by the authors.

Figure 1
Figure 1. An example of binary patterns used to encode information on the image. The lattice of 3x3 elements contains 2 9 =512 such kernels. For encoding graphic in￾formation we use rectangular kernels of various sizes as shown in fig. 2. С = 385 С = 56 С = 274 С = 273 С = 84 С = 466 С = 79 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. a) 3D image of the network formed as a result of training; b) example of superposition of a rectangular 3x3 lattice within the aperture; c) scheme of a strong classifier. Unlike the coding scheme in [5], the used coding scheme (1) does not allow image filtering. It transforms the image into a feature space with a dimension much larger than the size of the original image, as in the Viola and Jones scheme. We will cal… view at source ↗
Figure 3
Figure 3. fig.3 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Formation of code description of a rectangular section of binary image: a) the considered image section with regions of diverse noises marked red and of the useful signal marked blue; b) the region of detector aperture showing several variants of non-local binary patte…
Figure 4
Figure 4. Figure 4: Study of the size of neural network for detecting railway car number depending on overlapping threshold and type of features [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The dependence of the FRR of a neural network for detecting railway car numbers on overlapping degree and type of feature used. FRR analysis shows that the detectors constructed using the Haar features are signif￾icantly inferior to the detectors based on binary templa…
Figure 6
Figure 6. Figure 6: Study of the size of the neural network for detecting digits on the railway car number depending on the overlapping threshold and type of the used features [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The dependence of the FRR of neural network for detecting digits on the numbers of railway cars on the degree of overlapping and type of the used feature. 5. Сonclusion The development of efficient detectors for character information recognition on an example of recogn…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 21 canonical work pages

  1. [36]

    https://github.com/telnykha/trains_dataset/

    work page

  2. [1]

    Rapid object detection using a boosted cascade of simple fe a- tures // IEEE Conf

    Viola P., Jones M.J. Rapid object detection using a boosted cascade of simple fe a- tures // IEEE Conf. on Computer V ision and Pattern Recognition. – Kauai, Hawaii, USA – 2001. V. 1. P. 511–518

    work page 2001

  3. [2]

    and Jones, M

    Viola, P. and Jones, M. Robust Real Time Object Detection, In International Jou r- nal of Computer Vision, 1(2), 2002

    work page 2002

  4. [3]

    (2013) Explaining AdaBoost

    Schapire R.E. (2013) Explaining AdaBoost. In: Schölkopf B., Luo Z., Vovk V. (eds) Empirical Inference. Springer, Berlin, Heidelberg

    work page 2013

  5. [4]

    Han, and Hanseok Ko Reinforced AdaBoost Learning for Object Detection with Lo cal Pattern Representations, Scientific World Journal

    Younghyun Lee, David K. Han, and Hanseok Ko Reinforced AdaBoost Learning for Object Detection with Lo cal Pattern Representations, Scientific World Journal. Sci- entific World Journal. Published online 2013 Nov 28. doi: 10.1155/2013/153465

    work page doi:10.1155/2013/153465 2013

  6. [5]

    Face Detection with the Modi fied Census Transform // Pr o- ceedings of the Sixth IEEE International Conference on Automatic Face and Ge sture Recognition (FGR’04) 0-7695-2122-3/04

    Froba B., Ernst A. Face Detection with the Modi fied Census Transform // Pr o- ceedings of the Sixth IEEE International Conference on Automatic Face and Ge sture Recognition (FGR’04) 0-7695-2122-3/04. 2004. IEEE

    work page 2004

  7. [6]

    Cheung, Yiu-ming & Deng, Junping. (2014). Ultra local binary pattern for image texture analysis. Proceedings 2014 IEEE International Conference on Security, Pattern Analysis, and Cybernetics, SPAC 2014. 290 -293. 10.1109/SPAC.2014.6982701. Jou r- nal of Network Communications and Emerging Technologies (JNCET) www.jncet.org Volume 7, Issue 5, May (2017) IS...

    work page arXiv 2014

  8. [7]

    Xie Liping, W. H. (July 2014). Video-based Facial Expression Recognition Using Histogram Sequence of Local Gabor Binary Patterns from Three Orthogonal Planes. Proceedings of the 33rd Chinese Control Conference. Nanjing, China

    work page 2014

Show all 23 references
  1. [8]

    Video analytics: Myths and reality // Security Focus

    Torsten A., Ivo K., Harald L. Video analytics: Myths and reality // Security Focus

  2. [9]

    Principles of pattern recognition

    Tu, D., Gonzalez R. Principles of pattern recognition. // M .: Tekhnosfera, 2012, ISBN: 978-5-94836-331-8. In Rusian. 10.Fukushima Kunihiko Neocognitron: A Self -organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position. // Biolog i...

    2012

  3. [18]

    Optical character recognition: a review

    Singh AK, Gupta A, Saxena A. Optical character recognition: a review. // Journal of Emerging Technolo gies and Innovative Research (JETIR) www.jetir.org 2016;3(4):142-146. 19.Mamta Kadyan A Review on Character Recognition Using OCR Algorithm // Journal of Network Communication...

    2017

  4. [20]

    Ankit Lat, C. V. Jawahar : Enhancing OCR Accuracy with Super Resol u- tion. ICPR 2018: 3162-3167

    2018

  5. [21]

    Chao Dong, Ximei Zhu, Yubin Deng, Chen Change Loy, and Yu Qiao Member, Boosting Optical Character Recognition: A Super -Resolution Approach, ICDAR 2015 COMPETITION ON TEXT IMAGE SUPER-RESOLUTION

    2015

  6. [22]

    Beketov, S.L.Karateev, Yu.V.Vizilter

    I.V. Beketov, S.L.Karateev, Yu.V.Vizilter. Algorithm for automatic d etection of license plate images of railway rolling stock based on adaptive boosting methods // Pr o- ceedings of the GraphiCon’2013 conference. Vladivostok. September 16 -20, 2013. Pp.219-221. In Russian

    2013

  7. [23]

    LeNet -5, convolutional neural networks

    LeCun, Yann. "LeNet -5, convolutional neural networks ". – http://yann.lecun.com/exdb/lenet/ 24.Xie L., Ahmad T., Jin L., Zang S. A new CNN Based method for multi - directional car licence plate detection /IEEE Trans. Intell. Transp. Syst 19, p.507 -517 (2018)

    2018

  8. [25]

    Ortiz License Plate Detection and Recognition Using Deeply Learned Convolutional Neural Networks https://arxiv.org/abs/1703.07330 (2017) 26.Scott Gray, Alec Radford and Diederik P

    Syed Zain Masood , Guang Shu , Afshin Dehghan , Enrique G. Ortiz License Plate Detection and Recognition Using Deeply Learned Convolutional Neural Networks https://arxiv.org/abs/1703.07330 (2017) 26.Scott Gray, Alec Radford and Diederik P. Kingma GPU Ke rnels for Block - Spars...

    2017 arXiv

  9. [29]

    and Yakhno V

    Bellustin N., Kalafati Y., Kovalchuck A., Telnykh A., Shemagina O., Sharma P., Vaish A., Verma S. and Yakhno V. Instant Human Face Attributes Recogn ition System 14 // International Journal of Advanced Computer Science and Applications (IJACSA ). Special Issue on Artificial In...

    2011

  10. [30]

    image s for adults

    N.S, Bellustin, Yu.D., Kalafati, A.A., Telnykh, O.V. Shemagina. Application of algorithms for adaptive segmentation and semantic description of images in the problem of recognition of “image s for adults” // Information -measuring and control systems

  11. [31]

    N. S. Belliustin, Yu. D. Kalafati, A. A. Telnykh, and O. V. Shemagina Neuron Like Algorithms of Adaptive Segmentation and Semantic Descri ption of Images in the ―Adult Image Recognition Problem // Optical Memory and Neural Networks (Info r- mation Optics). 2014. V. 23 (1). P. ...

    2014

  12. [34]

    Microcolumns in the cerebral cortex

    Jones E.G. Microcolumns in the cerebral cortex. // Proc. Natl. Acad. Sci. U. S. A. UNITED STATES, 2000. Vol. 97, № 10. P. 5019–5021

    2000

  13. [35]

    http://caffe.berkeleyvision.org

  14. [2013]

    Vol.11 No. 7. Pp. 37 - 42. In Russian

  15. [2019]

    P. 186. ISBN: 978-5-9901176-5-5. In Rusian

Pith tools

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