Pith. sign in

REVIEW 4 major objections 5 minor 19 references

A Novel Convolutional Neural Network-Based Framework for Complex Multiclass Brassica Seed Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A custom 23-layer CNN trained from scratch reaches 93 percent accuracy on ten visually similar Brassica seed classes, and the paper argues that beats DenseNet121, InceptionV3, and ResNet152.

desk verdict A new Brassica seed dataset is the real contribution, but the headline 93% accuracy is invalid because the test set was used to pick hyperparameters. read the letter →

arxiv 2505.21558 v1 pith:ENUBKWZY submitted 2025-05-26 cs.CV cs.AIcs.LG

Elhoucine Elfatimia , Recep Eryigitb , Lahcen Elfatimi This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords brassicaseedclassificationconvolutionalneuralnetworkimagedatasetmulticlasstexture-similarityagriculturaldeeplearningpretrainedmodelcomparison
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 a bespoke 23-layer convolutional network, built with alternating $5\times5$ and $3\times3$ filters and trained from scratch, classifies ten texture-similar Brassica seed types at 93 percent test accuracy on a newly collected dataset of 6,065 microscope images. The claim matters because seed-type identification underpins purity control, yield estimation, and early risk assessment in farming, and because the task is genuinely hard: the seeds look alike and some types arise from crossing. The paper's central assertion is that a small purpose-built network exceeds standard large architectures on this crop, reporting DenseNet121 at 90.03 percent, InceptionV3 at 84.71 percent, and ResNet152 at 73.34 percent.

What carries the argument

The load-bearing object is the proposed 23-layer CNN itself: an input layer at $128\times128$ pixels, convolutional layers alternating $5\times5$ and $3\times3$ filters to capture features at different scales, pooling layers (with the paper also describing identity-mapping layers in place of downsampling), and a head of two dense layers of 512 neurons ending in a ten-way softmax. It is trained from scratch, not fine-tuned, with batch size 64, learning rate 0.001, Adam, and categorical cross-entropy for 200 epochs. The supporting object is the new ten-class Brassica dataset of 6,065 images captured with a digital microscope at $1600\times1200$ pixels and resized to $128\times128$, used to benchmark the proposed model against three pretrained networks under a shared evaluation protocol.

What would settle it

Retrain the proposed model and DenseNet121 with every hyperparameter fixed using only the validation set, across several random initializations, and compare test accuracy; if the gap nearly vanishes, the 93 percent result reflects test-set tuning rather than a property of the architecture.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a comparatively small 23-layer CNN, using stacked convolutional layers with $5\times5$ and $3\times3$ filters whose output spatial dimensions shrink from $42\times42$ down to $2\times2$, can recognize ten Brassica seed classes with 93 percent test accuracy, 95.56 percent training accuracy, and 94.21 percent validation accuracy after 200 epochs with a batch size of 64, the Adam optimizer, and a learning rate of 0.001. The architecture ends in two dense layers of 512 neurons and a softmax head, and the authors attribute its strength to the mixture of filter sizes capturing fine spatial detail that pretrained networks miss. The result is reported on a new dataset of 6,065 digital-microscope images split 50/30/20 into training, testing, and validation, and the comparison against DenseNet121, InceptionV3, and ResNet152 is interpreted as showing that the proposed model is significantly better, with the only notable confusions occurring between two texture-similar classes.

Load-bearing premise

The 93 percent figure requires that the test set never influenced model choices, but the paper says batch size 64 and 200 epochs were picked because they maximized testing accuracy, so that independence is not met and the reported number is at best an optimistic estimate.

Editorial extensions

If this is right

  • A purpose-built network with varied filter sizes can outperform much larger pretrained models on a domain with subtle texture differences, suggesting that architecture choice matters more than raw scale for such tasks.
  • The collected Brassica dataset gives the field concrete reference numbers: future seed-classification models can be compared against the reported 93 percent, 90.03 percent, 84.71 percent, and 73.34 percent figures.
  • The per-class analysis shows the remaining errors concentrate in two texture-similar classes, which points to where better lighting control or additional data would add the most value.
  • Because the proposed model is smaller and trains faster than the pretrained baselines, the approach could make high-accuracy seed sorting feasible on modest hardware, which matters for on-farm use.
  • The authors state that the framework can be extended to other visually similar seed and object classification problems, not just Brassica.

Reading between the lines

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

  • Because Section 4.1 reports choosing batch size 64 and 200 epochs from the testing-accuracy curves, the 93 percent figure is likely optimistic; a rerun with hyperparameters fixed on the validation set alone would probably narrow the gap to DenseNet121.
  • The architecture description contradicts itself, saying MaxPooling2D layers are replaced by identity mappings while also listing spatial dimensions that shrink, and the paper reports two different training/validation accuracy pairs (96.10/95.47 and 95.56/94.21) for the same model, so the exact configuration behind the 93 percent result is not fully pinned down.
  • The superiority claim rests on a single training run per architecture; several runs with different random initializations would show whether the 2.97-percentage-point edge over DenseNet121 is reproducible.
  • The authors attribute the main errors, classes 5 and 7, to texture similarity and camera lighting; photographing those two classes under controlled illumination would give a direct test of where the remaining error comes from.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a 23-layer convolutional neural network for classifying ten morphologically similar Brassica seed types from a newly collected image dataset. The authors report a 93% test accuracy for their architecture and claim it is significantly better than three pretrained baselines: DenseNet121 (90.03%), InceptionV3 (84.71%), and ResNet152 (73.34%). The manuscript also describes dataset construction, hyperparameter choices, and per-class performance metrics, and it concludes that the custom CNN combines depth and width to achieve superior generalization on this task.

Significance. If the reported 93% test accuracy were a valid, unbiased out-of-sample estimate, the paper would constitute a useful empirical contribution: a compact, custom architecture beating large pretrained models on a fine-grained, domain-specific classification task, together with a new Brassica seed dataset. However, the evaluation protocol is compromised because hyperparameters were selected using the test set, and the manuscript contains multiple unresolved numerical and methodological contradictions. The central claim is therefore not established by the present evidence. The dataset collection effort and the comparative experiments are in principle valuable, but the current presentation does not permit a reader to verify or reproduce the results.

major comments (4)
  1. [§4.1, Figs. 5 and 6] The test set is used for model selection: Section 4.1 explicitly states that batch size 64 and 200 epochs were chosen because they maximize testing accuracy (Figs. 5b and 6). Consequently, the reported 93% test accuracy is the maximum over the explored hyperparameter grid, not an independent holdout estimate. This biases the headline number and invalidates the comparison with the baselines in Table 3, since the same test set was used to choose the settings that are then reported as superior. The authors must re-run the evaluation with a held-out test set that is never used for any model or hyperparameter selection, or report results on a separate final test set.
  2. [§3.1, Table 2, §4.2.1] The dataset size is internally inconsistent. The text in Section 3.1 states that a total of 6065 images were divided into ten classes, but the numbers in Table 2 sum to 5925 images. Moreover, Section 4.2.1 states that the test set contained 1214 images, which is neither 30% of 6065 nor 30% of 5925. These discrepancies make it impossible to determine the actual dataset split or to reproduce the reported accuracy. The authors need to correct the counts and reconcile the stated split percentages with the actual number of images in each partition.
  3. [§3.3, §4.3] The description of the baseline models is self-contradictory: Section 3.3 says the pretrained models used 'pre-trained weights that were randomly initialized during the training process,' and Section 4.3 refers to 'updated weights and fine-tuning.' Randomly initialized weights are not pretrained weights. If the baselines were trained from scratch on roughly six thousand images, then the comparison is not against pretrained transfer-learning models, and the claim that a small custom CNN outperforms pretrained state-of-the-art architectures is misleading. The authors must clarify the exact initialization and training protocol for each baseline.
  4. [§4.2 vs. §4.2.1] The reported accuracy values are contradictory between sections. Section 4.2 reports an average training accuracy of 96.10% and a validation accuracy of 95.47%, while Section 4.2.1 reports a training accuracy of 95.56% and a validation accuracy of 94.21% after 200 epochs, alongside the 93% test accuracy. These numbers cannot all describe the same final model. The authors must provide a single, consistent set of final training, validation, and test metrics, ideally together with the standard deviations over multiple runs.
minor comments (5)
  1. [§3.3] The architecture description first states that 'convolutional layers ... are followed by max pooling layers' and later says that 'MaxPooling2D layers [are replaced] with identity mappings'; these statements are incompatible and should be reconciled with an exact layer-by-layer specification.
  2. [Eq. (3)] The F1-score formula is missing the factor of 2: the standard definition is F1 = 2 * Precision * Recall / (Precision + Recall), not Precision * Recall / (Precision + Recall). This should be corrected, and the reported F1 values should be checked against the corrected formula.
  3. [Throughout] The manuscript contains numerous typographical and grammatical errors, including 'classifiying,' 'Therfore,' 'algorithim,' 'Nave Bayes,' 'Comparaison,' and 'metrices,' as well as inconsistent spelling of DenseNet121 as 'Densnet121' and 'Densent121.' A careful language edit is needed.
  4. [§4.3 and Table 3] The claim that the proposed model is 'significantly better' than DenseNet121 is not supported by any statistical test, confidence interval, or repeated-run variability; a single run with a 2.97 percentage-point gap (93.0% vs. 90.03%) may not be statistically significant, especially given the test-set contamination.
  5. [General] The manuscript does not state whether the dataset, code, or trained models are publicly available, which hinders reproducibility; the authors should provide an availability statement or upload the data and code to a public repository.

Circularity Check

1 steps flagged · score 6.0 of 10

Test-set contamination: batch size 64 and 200 epochs were chosen because they maximized test-set accuracy (Section 4.1), so the reported 93% is a selection-biased estimate, not an independent holdout prediction.

  1. fitted input called prediction [Section 4.1 (Figs. 5 and 6); reported in Section 4.2.1 and Table 3, Section 4.3]
    "Further analysis revealed that a batch size of 64 yielded the most effective outcomes throughout the model training process. The testing accuracies at various model training epochs were evaluated and illustrated in Fig. 6. Notably, the testing accuracy displayed a gradual improvement with the progression of epochs, ultimately reaching a peak at 200 iterations. Therefore, 200 epochs were chosen as a balance between training time and accuracy."

    Hyperparameters are the fitted inputs: Section 4.1 reports that batch size 64 'yielded the most effective outcomes' and 200 epochs were chosen because test accuracy peaked at 200 iterations, i.e., both settings were selected by maximizing the test-set score. Section 4.2.1 then presents 'an accuracy of 93% on the test dataset' as the model's measured performance, and Section 4.3 uses the same number to claim superiority over DenseNet121 at 90.03%. Once the test split has been used to choose batch size and epoch count, the 93% is not an independent prediction on unseen data; it is the best point of the tested hyperparameter grid.

full rationale

The paper makes no formal derivation and cites no self-authored uniqueness theorem, so the self-citation categories do not apply. The circularity is in the evaluation protocol. Section 4.1 explicitly uses the test split to choose hyperparameters: batch size 64 is adopted because testing accuracy 'soared' with batch size, and 200 epochs because test accuracy reached a peak at 200 iterations. The same test split then produces the headline 'accuracy of 93% on the test dataset' (Section 4.2.1) and the comparison with DenseNet121 in Table 3. Consequently, the reported 93% is a maximum over the explored batch-size/epoch grid rather than an unbiased out-of-sample estimate, and the claimed two-plus-point advantage over DenseNet121 is not established by a clean held-out experiment. The architecture description itself is self-contained and the performance metrics are standard; the flaw is that the central number is partly fitted to the test set. The score of 6 reflects that the headline prediction reduces to test-set selection, while the rest of the manuscript is ordinary empirical work rather than a definitional derivation.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim (93% test accuracy) rests on a set of tuned hyperparameters and on the dataset being correctly labeled. No external benchmark or independent dataset is used. The paper's main quantitative claim is a fitted performance value rather than a derived prediction.

free parameters (4)
  • Batch size = 64
    Selected after comparing testing accuracy across batch sizes 8, 16, 32, and 64 in Section 4.1; the reported final accuracy depends on this choice.
  • Number of epochs = 200
    Chosen because testing accuracy peaked at 200 iterations (Fig. 6); no later epochs were tested.
  • Learning rate = 0.001
    Set as 'appropriate' in Sections 3.2 and 4.2; no search or sensitivity analysis is shown.
  • Architecture hyperparameters = 23 layers; 5x5 and 3x3 filters; 512-unit dense layers
    Hand-designed and adjusted to improve accuracy on the collected dataset; no evidence of optimality beyond the reported single run.
assumptions (3)
  • domain assumption The collected images are correctly labeled and each of the ten classes is homogeneous.
    No expert verification, inter-labeler agreement, or quality control is reported; the classification accuracy is meaningless if labels are wrong.
  • domain assumption The test set was held out and not used for model or hyperparameter selection.
    Section 4.1 selects batch size and epochs based on testing accuracy, violating this assumption.
  • domain assumption The described architecture can be implemented as stated.
    The text contradicts itself about max pooling vs. identity mappings (Section 3.3), so the exact network is ambiguous.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Convolutional Neural Network-Based Framework for Complex Multiclass Brassica Seed Classification." pith.science (2026). https://pith.science/paper/ENUBKWZY

@misc{pith2026250521558,
  author       = {Pith},
  title        = {Pith review of: A Novel Convolutional Neural Network-Based Framework for Complex Multiclass Brassica Seed Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ENUBKWZY}},
  note         = {Machine review of arXiv:2505.21558}
}
read the original abstract

Agricultural research has accelerated in recent years, yet farmers often lack the time and resources for on-farm research due to the demands of crop production and farm operations. Seed classification offers valuable insights into quality control, production efficiency, and impurity detection. Early identification of seed types is critical to reducing the cost and risk associated with field emergence, which can lead to yield losses or disruptions in downstream processes like harvesting. Seed sampling supports growers in monitoring and managing seed quality, improving precision in determining seed purity levels, guiding management adjustments, and enhancing yield estimations. This study proposes a novel convolutional neural network (CNN)-based framework for the efficient classification of ten common Brassica seed types. The approach addresses the inherent challenge of texture similarity in seed images using a custom-designed CNN architecture. The model's performance was evaluated against several pre-trained state-of-the-art architectures, with adjustments to layer configurations for optimized classification. Experimental results using our collected Brassica seed dataset demonstrate that the proposed model achieved a high accuracy rate of 93 percent.

Figures

Figures reproduced from arXiv: 2505.21558 by the authors.

Figure 5
Figure 5. The effect of batch sizes on the model’s performance is shown in (a) batch size vs. training time per epoch and (b) batch size vs. the model’s testing accuracy [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 11
Figure 11. Comparative analysis of classification performance based on performance metrices: a) recall, b) precision, c) f1 score. performance. Next, in addition to the model accuracy, we used class accu￾racy, which seems more descriptive. We analyzed the performance of each seed class mentioned in the proposed model in terms of precision, recall, and F1-score, as shown in [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

  1. [1]

    Universit., seed collection and plant genetic diversity, https://embryo.asu.edu/pages/seed-collection-and-plant-genetic- diversity-1900-1979

    A.S. Universit., seed collection and plant genetic diversity, https://embryo.asu.edu/pages/seed-collection-and-plant-genetic- diversity-1900-1979

    work page 1900

  2. [2]

    Agrawal, P

    D. Agrawal, P. Dahiya, Comparisons of classification algorithms on seeds dataset using machine learning algorithm, Compusoft 7 (5) (2018) 2760–2765

    work page 2018

  3. [3]

    Foysal, M

    F.A. Foysal, M. Shakirul Islam, S. Abujar, S. Akhter Hossain, et al., A novel approach for tomato diseases classification based on deep convolutional neural networks, in: Proceedings of International Joint Conference on Computational Intelligence, Springer, 2020, pp. 583–591. URL: https://doi.org/10.1007/978-981-13-7564-4 _49

    work page doi:10.1007/978-981-13-7564-4 2020

  4. [4]

    Gulzar, Y

    Y. Gulzar, Y. Hamid, A.B. Soomro, A.A. Alwan, L. Journaux, A convolution neural network -based seed classification system, Symmetry 12 (12) (2020) 2018. URL: https://www.mdpi.com /2073-8994/12/12/2018

    work page 2020

  5. [5]

    M. Keya, B. Majumdar, M.S. Islam, A robust deep learning segmentation and identification approach of different bangladeshi plant seeds using cnn, in: 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), IEEE, 2020, pp. 1 –6. URL: 10.1109/ICCCNT49239. 2020.9225677

    work page arXiv 2020

  6. [6]

    A. Ali, S. Qadri, W.K. Mashwani, S. Brahim Belhaouari, S. Naeem, S. Rafique, F. Jamal, C. Chesneau, S. Anam, Machine learning approach for the classification of corn seed using hybrid features, International Journal of Food Properties 23 (1) (2020) 1110 –1124. URL: 10.1080/10942912.2020.1778724

    work page arXiv 2020

  7. [7]

    Salimi, B

    Z. Salimi, B. Boelt, Classification of processing damage in sugar beet (beta vulgaris) seeds by multispectral image analysis, Sensors 19 (10) (2019) 2360. URL: https://doi.org/10.3390/s19102360

    work page doi:10.3390/s19102360 2019

  8. [8]

    Jung, J.S

    M. Jung, J.S. Song, S. Hong, S. Kim, S. Go, Y.P. Lim, J. Park, S.G. Park, Y.M. Kim, Deep learning algorithms correctly classify brassica rapa varieties using digital images, Frontiers in Plant Science 12. URL: DOI:10.3389/fpls.2021.738685

    work page arXiv

Show all 19 references
  1. [9]

    Dubey, S

    B. Dubey, S. Bhagwat, S. Shouche, J. Sainis, Potential of artificial neural networks in varietal identification using morphometry of wheat grains, Biosystems engineering 95 (1) (2006) 61 –67. URL: DOI:10.1016/j.biosystemseng.2006.06.001

    2006 doi

  2. [10]

    Guevara -Hernandez, J.G

    F. Guevara -Hernandez, J.G. Gil, A machine vision system for classification of wheat and barley grain kernels, Spanish Journal of Agricultural Research (3) (2011) 672–680. URL: DOI:10.5424/ sjar/20110903-140-10

    2011

  3. [11]

    Shahid, M.S

    M. Shahid, M.S. Naweed, S. Qadri, Mutiullah, E.A. Rehmani, Varietal discrimination of wheat seeds by machine vision approach, Life Sci 11 (6) (2014) 245–252

    2014

  4. [12]

    Loddo, M

    A. Loddo, M. Loddo, C. Di Ruberto, A novel deep learning based approach for seed image classification and retrieval, Computers and Electronics in Agriculture 187 (2021) 106269. URL: https: //doi.org/10.1016/j.compag.2021.106269

    2021

  5. [14]

    Szegedy, W

    C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S.E. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014) 1–9. URL: https://doi.org/ 10.48550/arXiv.1409.4842

    2014 doi arXiv

  6. [15]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural networks, Communications of the ACM 60 (6) (2017) 84–90. URL: DOI:10.3390/app10041245

    2017 doi

  7. [16]

    Szegedy, V

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 2818 –2826. URL: https://doi.org/10.48550/a rXiv.1512.00567

    2015 doi arXiv

  8. [17]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 770–778. URL: DOI:10.1109/CVPR.2 016.907

    2015 doi

  9. [18]

    Abed, A.A

    S. Abed, A.A. Esmaeel, A novel approach to classify and detect bean diseases based on image processing, 2018 IEEE Symposium on Computer Applications & Industrial Electronics (ISCAIE) (2018) 297–302. URL: DOI:10.1109/ISCAIE.2018.8405488

    2018

  10. [19]

    Powers, Evaluation: from precision, recall and f - measure to roc, informedness, markedness and correlation, ArXiv abs/2010.16061

    D.M.W. Powers, Evaluation: from precision, recall and f - measure to roc, informedness, markedness and correlation, ArXiv abs/2010.16061. URL: https://doi.org/10.48550/arXiv.2010.16 061

    2010 doi arXiv

  11. [20]

    10.1007/978-3-030-1 2388-8_50

    H. Alatrista -Salas, J. Morzán -Samamé, M.N. del Prado, Crime alert! crime typification in news based on text mining, Springer Verlag, Alemania, 2020, pp. 725 –741. URL: doi="10.1007/978-3-030-1 2388-8_50"

    2020 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.