Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Involution-Infused DenseNet with Two-Step Compression for Resource-Efficient Plant Disease Classification

T0 review · 4 major / 6 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 The paper claims a 0.29M-parameter involution-infused DenseNet keeps 96.99% and 98.63% leaf-disease accuracy after two-step compression, and stays precision-recall balanced under blur where larger compressed students collapse.

desk verdict Incremental application of the authors' own compression pipeline and involution layers, with a plausible efficiency story but unreliable reported numbers and test-set-tuned pruning rates. read the letter →

arxiv 2506.00735 v1 pith:CECHK2IS submitted 2025-05-31 cs.CV

T. Ahmed , S. Jannat , Md. F. Islam , J. Noor This is my paper
classification cs.CV
keywords PlantLeafDiseasesConvolutionalNeuralNetworkInvolutionProcessWeightPruningKnowledgeDistillationDenseNetmodelcompressionnoiserobustness
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 argues that plant-disease classifiers can be shrunk to phone-sized models without losing accuracy if three techniques are combined: distilling a large pretrained network into a small student, pruning the student's smallest weights, and mixing in involution layers whose filters are generated from the input image rather than fixed after training. The authors build Hybrid_DenseNet, a compressed DenseNet169 with three involution layers, and report 96.99% accuracy on the 38-class PlantVillage set and 98.63% on the four-class Paddy Leaf set after 10% pruning, with 0.29M parameters and about 0.745 billion FLOPs. Their central claim is robustness: under Gaussian blur the hybrid keeps balanced precision and recall (F1 0.82 on PlantVillage), while the plain distilled DenseNet collapses to 0.55 and two compressed students cannot be pruned at all. If the claim holds, real-time disease screening on edge devices is feasible at near-state-of-the-art accuracy, and the compression also lowers the energy cost of inference.

What carries the argument

The machinery is the two-step compression pipeline plus the involution layer. Knowledge distillation trains a compact student against temperature-scaled soft outputs of a pretrained teacher, blending a KL-divergence soft loss with ground-truth cross-entropy (alpha 0.5, temperature 3-5); post-training weight pruning then zeroes every weight below a percentile threshold with a binary mask, sparing biases, batch-norm, and the last layer. The involution layer — a kernel generated dynamically from the input at each position, using a reduction ratio of 2 — is inserted after each DenseBlock of the student. The load-bearing interaction is that dense feature reuse plus position-adaptive filtering keeps per-class sensitivity when images degrade, which the noise experiment is designed to expose.

What would settle it

Re-run the two-step pipeline with the pruning budget fixed before any test-set evaluation (chosen on a validation split), across several random seeds, reporting mean and standard deviation: if the hybrid's 96.99% and 98.63% margins over the pruned ResNet50 and the plain DenseNet student vanish or invert, the central claim fails. As a second check, blur a held-out test set at several Gaussian severities and verify that the involution hybrid keeps F1 above 0.8 while DenseNet169's stays collapsed; if both degrade together, the involution layers are not doing the claimed work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that infusing involution layers into a compressed DenseNet changes what survives both compression and image degradation. A DenseNet169-derived student with three DenseBlocks of [3,4,5] layers, growth rate 32, and an involution layer after each block reaches 97.10% accuracy on PlantVillage and 98.87% on Paddy Leaf after knowledge distillation, and a 10% magnitude-pruning pass holds 96.99% and 98.63% with 0.29M parameters and 0.745B FLOPs. The ablation shows accuracy rising as involution layers increase from one to three, and the noise test shows the hybrid keeping balanced precision and recall (0.88/0.81, F1 0.82 on PlantVillage) after Gaussian blur, where the plain DenseNet student's F1 falls to 0.55 and MobileNet-V2 and EfficientNet-B0 destabilize under further pruning. The authors conclude that dynamically generated involution kernels preserve class-wise spatial sensitivity when high-frequency cues are blurred, making the hybrid the most balanced model for resource-constrained deployment, even though ResNet50 keeps a slightly higher raw F1 under noise at more than twice the compute.

Load-bearing premise

The load-bearing premise is that the reported accuracies are honest optima: pruning rates were chosen by scanning 0% to 99% and keeping whichever rate scored best on the test set, so the headline numbers fit that test set rather than predict a fresh one, and a fixed a-priori rate or repeated runs with standard deviations could reorder the models.

Editorial extensions

If this is right

  • A 0.29M-parameter model with 0.745B FLOPs holds 96.99% accuracy on the 38-class PlantVillage set and 98.63% on Paddy Leaf, putting real-time leaf-disease screening within reach of phones and drones.
  • The same recipe compresses DenseNet169 from 14M parameters and 6.76B FLOPs to 0.257M parameters and 0.67B FLOPs while keeping 99.21% (Paddy Leaf) and 95.04% (PlantVillage) accuracy after two steps.
  • Adding involution layers raises accuracy monotonically in the ablation (95.91 to 96.76 to 97.10 on PlantVillage) for roughly 0.01-0.03M extra parameters per layer, until overfitting risk grows.
  • Under Gaussian blur the hybrid is the only compressed model with precision and recall both above 0.8 on PlantVillage; MobileNet-V2 and EfficientNet-B0 students destabilize when pruned after distillation, so the backbone limits how much compression any pipeline can apply.
  • FLOP reductions of about 89% for the hybrid versus its DenseNet169 teacher, and 96.7% for the VGG16 student, imply proportionally lower energy per inference, which the authors tie to sustainable deployment.

Reading between the lines

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

  • The headline numbers are selected rather than predicted: pruning rates were scanned from 0% to 99% and the rate with the best test-set accuracy was reported as optimal, so a fixed a-priori budget or validation-only selection could reorder the models and shrink the claimed margins.
  • The robustness claim is tested under a single degradation, Gaussian blur; if involution preserves spatial sensitivity as argued, the same advantage should appear under motion blur, low resolution, and JPEG compression, which the paper does not test.
  • The paper's own limitation note says minority classes score below 0.40, so the aggregate F1 hides serious per-class failures; class-balanced sampling or a focal loss would likely change both the accuracy and the noise numbers.
  • FLOPs are a proxy, not a measurement; reporting measured energy per image on a phone-class CPU would turn the efficiency argument into a deployable number.
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 / 6 minor

Summary. The paper proposes a two-step compression pipeline (knowledge distillation followed by post-training weight pruning) applied to five CNN baselines, and introduces a DenseNet169-based student model augmented with involution layers for plant disease classification on PlantVillage and PaddyLeaf. It reports post-compression accuracy, parameter counts, inference time, FLOPs, and a Gaussian-blur robustness test, and concludes that the proposed Hybrid_DenseNet is the most robust and efficient model after compression.

Significance. If the claims were substantiated, the work would be a useful engineering contribution for resource-constrained agricultural disease detection. However, the novelty is incremental: Section 5.2 explicitly builds on the authors' own previous two-step compression framework [31], and the new contribution is limited to inserting involution layers into the DenseNet student. The paper's empirical evidence is weakened by a test-set-based pruning-rate selection procedure, multiple internal numerical contradictions, and a robustness evaluation with no noise parameters or error bars. The paper does not provide code or trained models, so the reported comparisons are not independently reproducible.

major comments (4)
  1. [§5.2, Algorithm 1; §4] The optimal pruning percentage is selected by iteratively evaluating 0%–99% pruning on the evaluation dataloader (Algorithm 1, 'D: Dataloader (for evaluation)'), and Section 4 states that PlantVillage is split 80/20 into training and testing with no validation set. Consequently, the 'Optimal Pruning Rate' in Tables 2 and 5, and the headline accuracies in Table 5 and 6 (e.g., 96.99% for Hybrid DenseNet after 10% pruning), are the result of a test-set scan over roughly 100 candidate rates. This is a selection-on-the-test-set procedure that inflates reported accuracy and invalidates the comparison against the DenseNet student (95.04%) and other models. The authors should fix the pruning rate a priori or select it on a validation split, and report repeated-run means and standard deviations.
  2. [Abstract, §7, Table 5] The ResNet50 results are swapped in the abstract and conclusion: 99.55% and 98.99% are attributed to PlantVillage and PaddyLeaf, respectively, whereas Table 5 reports 99.55% on Paddy Leaf and 98.99% on PlantVillage. In addition, Section 7 states that the original DenseNet169 achieves 99.21% on Paddy Leaf and 93.56% on PlantVillage, and the abstract reports 93.96% for the DenseNet-based model, but Table 5 lists 99.21% on Paddy Leaf and 95.04% on PlantVillage. These mismatches mean the reader cannot rely on the reported accuracy numbers without correction.
  3. [§6.3, Table 6] The text states that the three-layer hybrid model on PlantVillage achieves 0.99 recall and F1 score, but Table 6 lists Recall 0.91 and F1 0.91 for exactly that configuration. Since the central claim of balancing accuracy and class-wise sensitivity depends on these metrics, this contradiction must be resolved.
  4. [§6.4, Table 7] The robustness comparison is based on a single application of Gaussian blur with no kernel size or sigma parameters, no repeated runs or error bars, and Step 2 entries for MobileNet_V2 and EfficientNet_B0 are marked N/A. Moreover, the hybrid does not have the highest accuracy in Table 7: on Paddy Leaf, ResNet50 Step 2 accuracy is 98.87% versus the proposed 98.76%. Thus the conclusion that 'Hybrid_DenseNet stands out as the most robust and efficient model' is not supported by the presented evidence.
minor comments (6)
  1. [Table 1] The table header 'PlantVillage & RiceLeaf' is confusing; the two dataset columns should be labeled separately, and the acronym 'RiceLeaf' should be made consistent with 'Paddy Leaf' used elsewhere.
  2. [Table 7] The column header 'MACS (10 9)' should read 'MACS (10^9)' for clarity.
  3. [§6.4] The Gaussian blur robustness test should specify the exact noise parameters (kernel size, sigma) and explain how the blur was applied to the test images.
  4. [General] No code, trained model checkpoints, or data splits are provided, which limits reproducibility; the authors should consider releasing these artifacts.
  5. [References] Reference [34] is a blog post; for a methods paper on pruning, a peer-reviewed source would be more appropriate.
  6. [§7] The prose contains several grammatical issues, e.g., 'if we look at the first step, we can see that the student model's training and learning curves...'; a careful language edit is advised.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported optimal accuracies are selected by scanning pruning rates on the test set, so the headline numbers are fitted maxima rather than independent predictions.

  1. fitted input called prediction [Section 4 (dataset split), Section 5.2 (Algorithm 1/pruning paragraph), Section 6.1-6.2 and Tables 2/5]
    "This pruning approach is iteratively evaluated across various pruning percentages (0%–99%) to determine the optimal trade-off between model efficiency and classification performance."

    For PlantVillage there is no validation split, so the pruning percentage p is selected by evaluating accuracy on the same test set that later generates the reported headline accuracy. Each model's 'Optimal Pruning Rate' is therefore the argmax of the test-accuracy scan, and the reported accuracy is the value of that same objective at the argmax: reported_acc = max_p Evaluate(prune(M,p), D_test). The headline numbers (e.g., Hybrid_DenseNet 96.99% at 10% pruning in Table 6) are thus statistically forced by the selection procedure, not independent predictions. The claimed superiority over the DenseNet student and the robustness ranking inherit this test-set selection. This is the fitted-input/called-prediction pattern: the evaluation data is reused to choose the reported operating point.

full rationale

The compression equations (Eqs 1-8) and Algorithm 1 are standard and fully specified, and the accuracy measurements are empirical, so there is no self-definitional equation-level circularity. The authors' self-citations ([31] for the two-step compression framework, [28] for involution) are not load-bearing because the current paper re-specifies the pipeline and provides its own ablation. The central circularity-like issue is the pruning-rate selection protocol: the 'optimal' rate is chosen by scanning 0-99% on the test set (no validation set for PlantVillage), and the same test set is then used to report the final accuracy. This makes the headline 'optimal' results selected maxima rather than predictive claims, which is a partial circularity in the evaluation of the central efficiency-robustness claim. A fixed or validation-selected pruning rate, plus variance estimates, would remove the issue.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claims rest on standard compression techniques, a public dataset benchmark, and several hand-chosen hyperparameters; no new theoretical entities are introduced.

free parameters (5)
  • alpha (knowledge distillation loss weight) = 0.5
    Chosen by hand; balances soft and hard losses in Eq. 7, Section 5.2.
  • temperature (logit scaling) = 3.00 to 5.00
    Chosen by hand for softening teacher outputs (Eq. 3-5).
  • optimal pruning rates per model and dataset = e.g., ResNet50 30% (Paddy), 25% (PlantVillage); DenseNet169 10% both; VGG16 30%/15%; MobileNetV2 20%/10%…
    Selected by scanning 0-99% pruning percentages against test accuracy (Section 5.2, Table 5).
  • student DenseNet architecture = 3 DenseBlocks with [3,4,5] layers, growth rate 32, channel width 16, 0.29-0.32M parameters
    Hand-designed reduction of DenseNet169 (Section 5.3).
  • number of involution layers = 3 (best of 1, 2, 3)
    Chosen via ablation against test accuracy (Section 6.3, Table 6).
assumptions (3)
  • domain assumption PlantVillage and PaddyLeaf datasets are representative of real-world plant disease images.
    Used for all accuracy claims; the authors note in Section 8 that PlantVillage images are clean with consistent backgrounds, so transfer to field conditions is uncertain.
  • domain assumption Gaussian blur is a representative degradation for robustness testing.
    Section 6.4 bases the robustness claim entirely on Gaussian blur noise.
  • domain assumption The teacher-student distillation setup preserves useful knowledge when the student architecture is heavily reduced.
    The entire compression pipeline assumes the hand-designed students can learn from the teachers; this is validated only by the reported test accuracies, not independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Involution-Infused DenseNet with Two-Step Compression for Resource-Efficient Plant Disease Classification." pith.science (2026). https://pith.science/paper/CECHK2IS

@misc{pith2026250600735,
  author       = {Pith},
  title        = {Pith review of: Involution-Infused DenseNet with Two-Step Compression for Resource-Efficient Plant Disease Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CECHK2IS}},
  note         = {Machine review of arXiv:2506.00735}
}
read the original abstract

Agriculture is vital for global food security, but crops are vulnerable to diseases that impact yield and quality. While Convolutional Neural Networks (CNNs) accurately classify plant diseases using leaf images, their high computational demands hinder their deployment in resource-constrained settings such as smartphones, edge devices, and real-time monitoring systems. This study proposes a two-step model compression approach integrating Weight Pruning and Knowledge Distillation, along with the hybridization of DenseNet with Involutional Layers. Pruning reduces model size and computational load, while distillation improves the smaller student models performance by transferring knowledge from a larger teacher network. The hybridization enhances the models ability to capture spatial features efficiently. These compressed models are suitable for real-time applications, promoting precision agriculture through rapid disease identification and crop management. The results demonstrate ResNet50s superior performance post-compression, achieving 99.55% and 98.99% accuracy on the PlantVillage and PaddyLeaf datasets, respectively. The DenseNet-based model, optimized for efficiency, recorded 99.21% and 93.96% accuracy with a minimal parameter count. Furthermore, the hybrid model achieved 98.87% and 97.10% accuracy, supporting the practical deployment of energy-efficient devices for timely disease intervention and sustainable farming practices.

Figures

Figures reproduced from arXiv: 2506.00735 by the authors.

Figure 1
Figure 1. Involution process [28] (a) PlantVillage (b) Paddy Leaf [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Sample images from the PlantVillage dataset [29] and the Paddy Leaf dataset [30] Ahmed et al.: Preprint submitted to Elsevier Page 15 of 14 [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Workflow of our study Ahmed et al.: Preprint submitted to Elsevier Page 16 of 14 [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: The architecture of the two-step compression pipeline Ahmed et al.: Preprint submitted to Elsevier Page 17 of 14 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: The proposed model architecture (a) Accuracy vs pruning percentage (b) Inference time vs pruning percentage [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Accuracy and inference time vs pruning percentage curves for the PlantVillage dataset Ahmed et al.: Preprint submitted to Elsevier Page 18 of 14 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Accuracy vs pruning percentage curves for the PlantVillage and Paddy Leaf datasets (a) Accuracy curve (b) Learning curve [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Accuracy and learning curves of the hybrid model for the PlantVillage dataset (a) Accuracy curve (b) Learning curve [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Accuracy curve and learning curves of the hybrid model for Paddy Leaf dataset Ahmed et al.: Preprint submitted to Elsevier Page 19 of 14 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Input image from PlantVillage (a) Layer 1 (b) Layer 2 (c) Layer 3 [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Involution kernel visualization of the hybrid Model for the PlantVillage dataset (a) VGG-16 base (b) VGG-16 (student) [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Accuracy curve of VGG-16 baseline and student model on Paddy Leaf dataset Ahmed et al.: Preprint submitted to Elsevier Page 20 of 14 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Accuracy curve of ResNet50 baseline and student model on PlantVillage dataset (a) Accuracy of light-weight models (b) Weighted parameter count of light-weight models [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Comparative analysis of the performance of the light-weight models Ahmed et al.: Preprint submitted to Elsevier Page 21 of 14 [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Comparative flop count of Teacher and Student models Ahmed et al.: Preprint submitted to Elsevier Page 22 of 14 [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 27 canonical work pages

  1. [31]

    Ahmed, S

    T. Ahmed, S. Jannat, A. Rahat, J. A. Mou, M. F. Islam, and J. Noor. Knowledge distillation and weight pruning for two-step compression of convnets in rice leaf disease classification. InProceedings of the 11th International Conference on Networking, Systems, and Security, NSysS ’24, pages 72–78. Association for Computing Machinery, New York, NY, USA, 2025...

    work page arXiv 2025

  2. [1]

    A.-Y. Dong, Z. Wang, J.-J. Huang, B.-A. Song, and G.-F. Hao. Bioinformatic tools support decision-making in plant disease management. Trends in Plant Science, 26(9):953–967, 2021.https://doi.org/10.1016/j.tplants.2021.05.001

    work page doi:10.1016/j.tplants.2021.05.001 2021

  3. [2]

    P. K. Krishnamurthy, R. J. Choularton, and P. Kareiva. Dealing with uncertainty in famine predictions: How complex events affect food security early warning skill in the greater horn of Africa.Global Food Security, 26:100374, 2020.https://doi.org/10.1016/j.gfs. 2020.100374

    work page arXiv 2020

  4. [3]

    Harish, D

    S. Harish, D. Saravanakumar, R. Radjacommare, E. Ebenezar, and K. Seetharaman. Use of plant extracts and biocontrol agents for the management of brown spot disease in rice.BioControl, 53:555–567, 2008.https://doi.org/10.1007/s10526-007-9098-9

    work page doi:10.1007/s10526-007-9098-9 2008

  5. [4]

    Barupal, M

    T. Barupal, M. Meena, and K. Sharma. A study on preventive effects of Lawsonia inermis L. bioformulations against leaf spot disease of maize.Biocatalysis and Agricultural Biotechnology, 23:101473, 2020.https://doi.org/10.1016/j.bcab.2019.101473

    work page arXiv 2020

  6. [5]

    Automaticidentificationofpeanut-leafdiseasesbasedonstackensemble.AppliedSciences,11(4):1950, 2021.https://doi.org/10.3390/app11041950

    H.Qi,Y.Liang,Q.Ding,andJ.Zou. Automaticidentificationofpeanut-leafdiseasesbasedonstackensemble.AppliedSciences,11(4):1950, 2021.https://doi.org/10.3390/app11041950

    work page doi:10.3390/app11041950 1950

  7. [6]

    Areviewofplantleaffungaldiseasesanditsenvironmentspeciation.Bioengineered,10(1):409– 424, 2019

    A.Jain,S.Sarsaiya,Q.Wu,Y.Lu,andJ.Shi. Areviewofplantleaffungaldiseasesanditsenvironmentspeciation.Bioengineered,10(1):409– 424, 2019

    work page 2019

  8. [7]

    K. P. Ferentinos. Deep learning models for plant disease detection and diagnosis.Computers and Electronics in Agriculture, 145:311–318, 2018.https://doi.org/10.1016/j.compag.2018.01.009

    doi:10.1016/j.compag.2018.01.009 2018

Show all 35 references
  1. [8]

    M.Donatelli,R.D.Magarey,S.Bregaglio,L.Willocquet,J.P.Whish,andS.Savary.Modellingtheimpactsofpestsanddiseasesonagricultural systems.Agricultural Systems, 155:213–224, 2017.https://doi.org/10.1016/j.agsy.2017.01.019

    2017 doi

  2. [9]

    Y. Li, H. Zhang, X. Xue, Y. Jiang, and Q. Shen. Deep learning for remote sensing image classification: A survey.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 8(6):e1264, 2018.https://doi.org/10.1002/widm.1264

    2018 doi

  3. [10]

    V. K. Sharma and R. N. Mir. A comprehensive and systematic look up into deep learning based object detection techniques: A review. Computer Science Review, 38:100301, 2020.https://doi.org/10.1016/j.cosrev.2020.100301

    2020

  4. [11]

    D. W. Otter, J. R. Medina, and J. K. Kalita. A survey of the usages of deep learning for natural language processing.IEEE Transactions on Neural Networks and Learning Systems, 32(2):604–624, 2020.https://doi.org/10.1109/TNNLS.2020.2979670

    2020

  5. [12]

    M. T. Ahad, Y. Li, B. Song, and T. Bhuiyan. Comparison of CNN-based deep learning architectures for rice diseases classification.Artificial Intelligence in Agriculture, 9:22–35, 2023.https://doi.org/10.1016/j.aiia.2023.07.001

    2023 doi

  6. [13]

    Prithviraj and S

    V. Prithviraj and S. Rajkumar. Magnitude-based weight-pruned automated convolutional neural network to detect and classify the plant disease. In A. B. Reddy, S. Nagini, V. E. Balas, and K. S. Raju, editors,Proceedings of Third International Conference on Advances in Computer E...

    2023

  7. [14]

    Q. Dong, R. Gu, S. Chen, and J. Zhu. Apple leaf disease diagnosis based on knowledge distillation and attention mechanism.IEEE Access, 12:65154–65165, 2024.https://doi.org/10.1109/ACCESS.2024.3397329

    2024

  8. [15]

    M. F. Islam, M. A. Manab, J. J. Mondal, S. Zabeen, F. B. Rahman, M. Z. Hasan, F. Sadeque, and J. Noor. Involution fused convolution for classifying eye-tracking patterns of children with autism spectrum disorder.Engineering Applications of Artificial Intelligence, 139:109475, ...

    2025

  9. [16]

    Macdonald, Y

    W. Macdonald, Y. A. Sari, and M. Pahlevani. Grow-light smart monitoring system leveraging lightweight deep learning for plant disease classification.Artificial Intelligence in Agriculture, 12:44–56, 2024.https://doi.org/10.1016/j.aiia.2024.03.003

    2024 doi

  10. [17]

    M.F.Islam,M.T.Reza,M.A.Manab,S.Zabeen,M.F.-U.Islam,M.F.Shahriar,M.Kaykobad,M.G.Z.A.Husna,andJ.Noor. Involution- based efficient autoencoder for denoising histopathological images with enhanced hybrid feature extraction.Computers in Biology and Medicine, 192:110174, 2025.https:...

    2025

  11. [18]

    Knowledgedistillationfacilitatesthelightweightandefficientplant diseases detection model.Plant Phenomics, 5:0062, 2023.https://doi.org/10.34133/plantphenomics.0062

    Q.Huang,X.Wu,Q.Wang,X.Dong,Y.Qin,X.Wu,Y.Gao,andG.Hao. Knowledgedistillationfacilitatesthelightweightandefficientplant diseases detection model.Plant Phenomics, 5:0062, 2023.https://doi.org/10.34133/plantphenomics.0062

    2023 doi

  12. [19]

    B. Hu, W. Jiang, J. Zeng, C. Cheng, and L. He. Fotca: hybrid transformer-cnn architecture using afno for accurate plant leaf disease image recognition.Frontiers in Plant Science, 14:1231903, 2023.https://doi.org/10.3389/fpls.2023.1231903

    2023

  13. [20]

    Akter and M

    R. Akter and M. I. Hosen. CNN-based leaf image classification for Bangladeshi medicinal plant recognition. In2020 Emerging Technology in Computing, Communication and Electronics (ETCCE), pages 1–6. IEEE, 2020.https://doi.org/10.1109/etcce51779.2020. 9350900

    2020

  14. [21]

    S. A. Riaz, S. Naz, and I. Razzak. Multipath deep shallow convolutional networks for large scale plant species identification in wild image. In2020InternationalJointConferenceonNeuralNetworks(IJCNN),pages1–7.IEEE,2020.https://doi.org/10.1109/ijcnn48605. 2020.9207113

    2020

  15. [22]

    M. S. R. Chy, M. R. H. Mahin, M. F. Islam, M. S. Hossain, and A. A. Rasel. Classifying corn leaf diseases using ensemble learning with dropout and stochastic depth based convolutional networks.Association for Computing Machinery, New York, NY, USA, 2023. https://doi.org/10.114...

    2023

  16. [23]

    Entropy-basedpruningmethodforconvolutionalneuralnetworks.TheJournalofSupercomputing,75:2950–2963,2019

    C.HurandS.Kang. Entropy-basedpruningmethodforconvolutionalneuralnetworks.TheJournalofSupercomputing,75:2950–2963,2019. https://doi.org/10.1007/s11227-018-2684-z

    2019 doi

  17. [24]

    E. C. Too, Y. Li, P. Kwao, S. Njuki, M. E. Mosomi, and J. Kibet. Deep pruned nets for efficient image-based plants disease classification. Journal of Intelligent & Fuzzy Systems, 37(3):4003–4019, 2019.https://doi.org/10.3233/jifs-190184

    2019 doi

  18. [25]

    S. W. Prakosa, J.-S. Leu, and Z.-H. Chen. Improving the accuracy of pruned network using knowledge distillation.Pattern Analysis and Applications, 24:819–830, 2021.https://doi.org/10.1007/s10044-020-00940-2

    2021 doi

  19. [26]

    J. J. Mondal, M. Islam, S. Zabeen, A. B. M. A. A. Islam, and J. Noor. Note: Plant leaf disease network (plead-net): Identifying plant leaf diseases through leveraging limited-resource deep convolutional neural network. 2022.https://doi.org/10.1145/3530190.3534844

    2022

  20. [27]

    Hinton, O

    G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network.arXiv, 2015.https://arxiv.org/abs/1503.02531

    2015 arXiv

  21. [28]

    M. F. Islam, S. Zabeen, M. A. Manab, M. R. H. Mahin, J. J. Mondal, M. T. Reza, M. Z. Hasan, M. Haque, F. Sadeque, and J. Noor. Med-ic: Fusing a single layer involution with convolUTIONS for enhanced medical image classification and segmentation.arXiv preprint arXiv:2409.18506, 2024

    2024 arXiv

  22. [29]

    Hughes, M

    D. Hughes, M. Salathé, and others. An open access repository of images on plant health to enable the development of mobile disease diagnostics.arXiv preprint arXiv:1511.08060, 2015

    2015 arXiv

  23. [30]

    P. K. Sethy, S. K. Behera, N. Kannan, S. Narayanan, and C. Pandey. Smart paddy field monitoring system using deep learning and IoT. Concurrent Engineering, 29(1):16–24, 2021.https://doi.org/10.1177/1063293X21988944

    2021 doi

  24. [32]

    Knowledgedistillationforefficientinstancesemanticsegmentationwithtransformers

    M.Li,M.Halstead,andC.Mccool. Knowledgedistillationforefficientinstancesemanticsegmentationwithtransformers. InProceedingsof the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 5432–5439, 2024

    2024

  25. [33]

    Bulinski and D

    A. Bulinski and D. Dimitrov. Statistical estimation of the Kullback–Leibler divergence.Mathematics, 9(5):544, 2021

    2021

  26. [34]

    M. Neo. A Comprehensive Guide to Neural Network Model Pruning.https://www.datature.io/blog/ a-comprehensive-guide-to-neural-network-model-pruning, [Accessed 18-09-2024]

    2024

  27. [35]

    Vente, L

    T. Vente, L. Wegmeth, A. Said, and J. Beel. From clicks to carbon: The environmental toll of recommender systems. InProceedings of the 18th ACM Conference on Recommender Systems, pages 580–590, 2024. Ahmed et al.:Preprint submitted to ElsevierPage 14 of 14 Involution-Infused D...

    2024

Pith tools

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