DragonFruitQualityNet: A Lightweight Convolutional Neural Network for Real-Time Dragon Fruit Quality Inspection on Mobile Devices

Md Robiul Islam; Md Zahurul Haquea; Muhammed Farhan Sadique Mahi; Syed Jubayer Jaman; Yeahyea Sarker

REVIEW 4 major objections 4 minor 18 references

DragonFruitQualityNet: A Lightweight Convolutional Neural Network for Real-Time Dragon Fruit Quality Inspection on Mobile Devices

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DragonFruitQualityNet is a compact CNN built to classify dragon fruit images into four quality grades and to run that classification on a phone, with a reported accuracy of 93.98%.

desk verdict The headline 93.98% is training accuracy; with 74.91% validation and no test set, the central claim is unsupported—though the mobile app integration is real engineering. read the letter →

arxiv 2508.07306 v1 pith:JOK5ZH5Q submitted 2025-08-10 cs.CV cs.AI

Md Zahurul Haquea , Yeahyea Sarker , Muhammed Farhan Sadique Mahi , Syed Jubayer Jaman , Md Robiul Islam This is my paper · ORCID

classification cs.CVcs.AI

keywords DragonFruitQualityNetlightweightCNNclassificationreal-timemobileinferencegradingpost-harvestmanagementagriculturalAIfour-classimage

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 sets out to show that a deliberately small convolutional neural network can sort dragon fruit images into four quality classes—fresh, immature, mature, and defective—well enough to be useful on a smartphone in the field. The authors assemble a 13,789-image public dataset, train their DragonFruitQualityNet architecture on 10,010 images, and report 93.98% accuracy as the model's headline performance. They also package the trained model into a mobile app so a farmer can photograph a fruit and receive a quality label without an internet connection. If the 93.98% figure describes performance on unseen images, the work provides a practical post-harvest grading tool for low-resource settings. The paper also reports that the same figure is the training accuracy and that validation accuracy reached 74.91%, so what the headline number measures is the key open question.

What carries the argument

The carrying object is the DragonFruitQualityNet architecture itself: a sequential CNN with repeated convolution-and-pooling blocks whose filter counts grow from 32 to 512, followed by dropout, flattening, a large fully connected layer, and a four-neuron softmax head. The companion mechanism is the mobile-deployment pipeline that converts the trained weights into a compact interpreter format and embeds them in a phone app for offline inference.

What would settle it

Take the trained DragonFruitQualityNet and run it on a held-out test set of dragon fruit images that were used neither for training nor for validation; compute accuracy per class. If the held-out accuracy lands near the 74.91% validation figure instead of 93.98%, the headline generalization claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that DragonFruitQualityNet—a CNN built from successive convolution blocks, max pooling, dropout, and a softmax output layer over four classes—achieves accuracy of 93.98% on dragon fruit quality classification while staying compact enough for real-time mobile deployment. The paper presents this as beating earlier fruit-quality classifiers and shows that inference can run on-device through a mobile app, with the trained model translated into a mobile-optimized format. The section reporting training details states that training accuracy reached 93.98% and validation accuracy 74.91%, which sets the boundary for what the paper has actually demonstrated.

Load-bearing premise

The central claim collapses if the 93.98% accuracy was measured on the training set rather than on unseen images; the paper reports that number as training accuracy and separately reports 74.91% validation accuracy, so a held-out test is the load-bearing premise.

Editorial extensions

If this is right

A farmer or inspector with a smartphone could photograph a dragon fruit and get a fresh, immature, mature, or defective label on the spot, without needing connectivity.
If the 93.98% number reflects generalization, DragonFruitQualityNet would beat earlier fruit-quality classifiers on the same kind of task and make automated grading accessible to smallholders at near-zero marginal cost.
The four-class, 13,789-image dataset with training-only augmentation provides a reusable benchmark for dragon fruit quality research.
Pre- and post-harvest workflows such as sorting, supply-chain monitoring, and market-grade decisions could be automated in low-connectivity environments.
The reported training setup—20 epochs on a modest model with dropout—suggests the result can be reproduced without large computational infrastructure.

Reading between the lines

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

The 74.91% validation accuracy in the training subsection suggests the 93.98% headline figure may describe fit to the training set rather than generalization; a held-out test evaluation is needed before deployment decisions.
Most of the model's 30.7M parameters sit in the final fully connected layer, which is unusual for a 'lightweight' design; replacing that layer with global average pooling or pruning it could cut size substantially, likely with little accuracy loss.
The same four-class pipeline and mobile packaging could transfer to other tropical fruits, such as mango or papaya, though the paper does not claim this.
A field trial measuring end-to-end latency and accuracy on actual phone cameras under varied lighting would settle whether the app is genuinely useful in real farming conditions; the paper's own limitations section says in-field performance still needs testing.

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.

Desk Editor's Note

read the letter

Hi, quick take on arXiv:2508.07306. The central claim is unsupported: the 93.98% accuracy in the title/abstract is training accuracy, not test or even validation. The paper's own §4.3.1 reports validation accuracy 74.91%. No held-out test set is ever described, and the confusion matrix described in §4.3 lists only 43 total misclassifications, which would give ~98.9% on validation or ~99.6% on training—reconcilable with neither reported number. That alone sinks the headline result.

What's actually decent: the authors assembled a 13,789-image dataset (drawing on Mendeley and their own captures), trained a standard VGG-style CNN with sensible augmentation and dropout, and built a Flutter app with a TFLite model, which is a real engineering effort. The app description is the one concrete artifact here.

The problems are proportionate. First, the headline metric is training fit, not generalization—a load-bearing error. Second, the 'lightweight' claim is contradicted by their own Table 2: 30.7M parameters, which is VGG-scale, not mobile-scale. Third, there are no baseline comparisons, so 'outperforms existing methods' is asserted, not demonstrated. Fourth, the data/code are not released, so nothing is reproducible. Fifth, some references look dubious (e.g., a 2025 Agriculture paper with a page number), so the citation list needs checking.

The introduction and literature review are fairly standard. The authors do cite prior dragon fruit work (YOLOv5, EfficientNet, Gencturk et al.), so they're aware this is an incremental application. The mobile app is novel as a deployment, but its performance is not measured—no latency, no field trials. The limitations section admits the dataset is homogeneous, which is honest but underscores the generalization gap.

For a reader: if you're interested in mobile agriculture apps, the Flutter/TFLite integration might be worth a skim. As a research result, it doesn't hold up. I would not cite it, and I would not spend referee time on it as-is. A desk reject is appropriate; the authors could resubmit if they actually evaluate on a held-out test set, add baselines, fix the confusion matrix numbers, and release code/data.

Verdict: reject.

Referee Report

4 major / 4 minor

Summary. The paper proposes DragonFruitQualityNet, a custom CNN for four-class dragon fruit quality classification (fresh, immature, mature, defective), trained on 13,789 images from a Mendeley dataset with a 10,010/3,779 training/validation split, data augmentation, and a TensorFlow Lite/Flutter mobile application. The abstract, introduction, and conclusion claim that the model achieves 93.98% accuracy and outperforms existing methods. Section 4.3.1, however, states that 93.98% is the training accuracy and that validation accuracy is 74.91%. No separate test set is described, and the confusion-matrix counts in Section 4.3 are inconsistent with both reported accuracies.

Significance. If the claimed 93.98% out-of-sample accuracy and real-time mobile performance were substantiated, the system could be practically useful for low-cost post-harvest grading. However, the manuscript provides no valid test-set evaluation, no baseline comparisons, no latency or throughput measurements, and no reproducibility artifacts. The only reported out-of-sample number is 74.91% validation accuracy, 19 points below the headline. The paper's own limitation statement in Section 5.1 admits the dataset is homogeneous and that in-field performance requires further testing, which undercuts the claimed 'practical validation.' As submitted, the core evidence is internally inconsistent and the central claims are unsupported.

major comments (4)

[Abstract; §4.3.1] The headline '93.98% accuracy' is explicitly a training accuracy in §4.3.1, while validation accuracy is 74.91%. Section 3.1 describes only a training/validation split; despite Section 3's mention of a test set, no test set is ever introduced. Presenting the training accuracy as the model's performance and as 'state-of-the-art accuracy' is circular: it is the objective value being fitted, not a generalization estimate. The central claim must be re-evaluated on a held-out test set.
[§4.3; Fig. 7] The confusion-matrix description lists 36 + 3 + 1 + 3 = 43 misclassified instances. For the validation set of 3,779 images this would imply 98.86% accuracy; for the training set of 10,010 images it would imply 99.57% accuracy. Neither value matches the reported 93.98% or 74.91%. The confusion matrix, as described, cannot be reconciled with either reported accuracy, so the experimental record is internally inconsistent and the reported metrics cannot be audited.
[§4.3; Section 2] The sentence 'DragonFruitQualityNet outperforms all contemporary deep learning models' is unsupported. No baseline model is trained, no quantitative comparison table is provided, and the literature review is narrative only. The abstract's claim of 'outperforming existing methods' therefore has no evidentiary basis and must be substantiated with controlled comparisons on the same data and protocol.
[§5.1; Contributions] Contribution 4 claims 'rigorous testing under realistic conditions,' but Section 5.1 states that the dataset is 'relatively homogeneous' and that 'in-field performance under variable conditions requires further testing.' These statements are contradictory. In addition, the term 'lightweight' is asserted without evidence: the model has approximately 30.7M parameters, dominated by a 19.66M-parameter dense layer, and no on-device inference latency, model-size benchmark, or comparison with MobileNet/EfficientNet is reported.

minor comments (4)

[Abstract; §3.1] The abstract says the dataset integrates 'self-collected samples with public datasets,' but Section 3.1 describes only a publicly available Mendeley dataset. Clarify the provenance.
[Equations (1)–(4)] The metric equations are garbled and incomplete (accents and fractions are corrupted). They should be typeset correctly so accuracy, precision, recall, and F1 are unambiguous.
[§4.3.1] The text mentions 'training and testing accuracy plots' and Figure 11, but no such plot is described in the text and the reference is vague. Please include the actual curves or remove the reference.
[§4.3.2; Table 3] The phrase 'Table 3[]' has an empty bracket; the table caption says 'Flutter dependencies and their uses' but the column header in the text reads 'purpose.' Minor formatting cleanup is needed.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline 93.98% accuracy is the training accuracy, not an out-of-sample evaluation; the only validation number reported is 74.91%, so the central performance claim reduces to the fitted training objective.

fitted input called prediction [Abstract; Section 4.3.1; Section 4.4]
"The proposed model achieves an impressive 93.98% accuracy, outperforming existing methods in fruit quality classification. ... Our proposed CNN model achieves an impressive 93.98% accuracy in correct class prediction, as demonstrated by the confusion matrix. ... the model achieved a training accuracy of 93.98%, while validation accuracy reached 74.91%."

The 93.98% figure is first presented as the model's achieved accuracy and as the basis for the 'outperforming existing methods' claim. Section 4.3.1 then states that this exact number is the training accuracy: 'the model achieved a training accuracy of 93.98%, while validation accuracy reached 74.91%.' Training accuracy is the metric the optimizer maximizes on the training split; reporting it as the model's classification performance turns the fitted objective value into the claimed result. Section 3.1 describes only a training/validation split (10,010/3,779) and never introduces a held-out test set, so no out-of-sample prediction is made at the 93.98% level. The only non-training number reported is 74.91% validation accuracy, which is materially lower. Thus the headline 'prediction' reduc

full rationale

The architecture, preprocessing, and training pipeline are standard and not circular: the model design is not claimed to be derived from the target accuracy. However, the paper's central quantitative claim is the 93.98% figure, and Section 4.3.1 explicitly identifies this figure as training accuracy. Presenting a training-set metric as the model's achieved accuracy means the headline is the value of the optimization objective, not an out-of-sample generalization result. The dataset section defines only training and validation splits; no test set is described, and the reported validation accuracy is 74.91%, far below the headline. Additionally, the text describing Fig. 7 lists only 43 misclassified instances (36+3+1+3), which cannot produce either 93.98% or 74.91% for the stated split sizes (10,010 training / 3,779 validation), making the reported evaluation internally inconsistent; this is a reproducibility/correctness problem rather than circularity. The paper's own limitation statement ('may not generalize across diverse lighting, fruit varieties, or device specifications') concedes that broader generalization remains untested. Because the headline accuracy reduces by construction to the training accuracy, the circularity score is 6.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on the correctness and representativeness of the public dataset labels, a random split with no leakage, and the validity of the training accuracy as a performance estimate. These are domain assumptions; no independent evidence is provided. Hyperparameters and augmentation ranges are hand-chosen with no sensitivity analysis.

free parameters (6)

Learning rate = 0.0001
Chosen manually as 'optimal' without a search or sensitivity analysis (§4.3.1).
Batch size = 32
Stated hyperparameter; affects convergence and generalization.
Number of epochs = 20
Training stopped at epoch 20; claimed convergence (§4.3.1).
Dropout rate = 0.5
Used after flatten and dense layers (§3.2).
Input image size = 256x256
Resized to match first conv2d layer (§3.2).
Augmentation ranges = rotation ±20°, zoom 15%, brightness/contrast ±10%
Chosen to simulate field variation; no ablation.

assumptions (4)

domain assumption The Mendeley dataset labels are correct and mutually exclusive.
The model's targets come from a public dataset without manual verification.
domain assumption The train/validation split is random and has no data leakage.
The paper does not describe the split mechanism.
domain assumption Validation accuracy is an unbiased estimate of field performance.
The paper presents validation results as evidence of real-world use.
standard math The TensorFlow/Keras implementation matches the described architecture.
No code or reproducibility artifacts are provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DragonFruitQualityNet: A Lightweight Convolutional Neural Network for Real-Time Dragon Fruit Quality Inspection on Mobile Devices." pith.science (2026). https://pith.science/paper/JOK5ZH5Q

@misc{pith2026250807306,
  author       = {Pith},
  title        = {Pith review of: DragonFruitQualityNet: A Lightweight Convolutional Neural Network for Real-Time Dragon Fruit Quality Inspection on Mobile Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JOK5ZH5Q}},
  note         = {Machine review of arXiv:2508.07306}
}

read the original abstract

Dragon fruit, renowned for its nutritional benefits and economic value, has experienced rising global demand due to its affordability and local availability. As dragon fruit cultivation expands, efficient pre- and post-harvest quality inspection has become essential for improving agricultural productivity and minimizing post-harvest losses. This study presents DragonFruitQualityNet, a lightweight Convolutional Neural Network (CNN) optimized for real-time quality assessment of dragon fruits on mobile devices. We curated a diverse dataset of 13,789 images, integrating self-collected samples with public datasets (dataset from Mendeley Data), and classified them into four categories: fresh, immature, mature, and defective fruits to ensure robust model training. The proposed model achieves an impressive 93.98% accuracy, outperforming existing methods in fruit quality classification. To facilitate practical adoption, we embedded the model into an intuitive mobile application, enabling farmers and agricultural stakeholders to conduct on-device, real-time quality inspections. This research provides an accurate, efficient, and scalable AI-driven solution for dragon fruit quality control, supporting digital agriculture and empowering smallholder farmers with accessible technology. By bridging the gap between research and real-world application, our work advances post-harvest management and promotes sustainable farming practices.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 14 canonical work pages

[1]

Zhang, Y

W. Zhang, Y. Cao, H. Zhang, and Z. Li, ―Light -weight Fruit -Detection Alg orithm for Edge Computing Applications,‖ Frontiers in Plant Science , vol. 12, Oct. 2021, Art. 740936. doi: 10.3389/fpls.2021.740936

work page arXiv 2021
[2]

Kamilaris and F

H. Kamilaris and F. X. Prenafeta -Boldú, ―Deep learning in agriculture: A survey,‖ Computers and Electronics in Agricult ure, vol. 147, pp. 70–90, Apr. 2018. doi: 10.1016/j.compag.2018.02.016

doi:10.1016/j.compag.2018.02.016 2018
[3]

M. A. Khan, M. Y. Javed, and R. Alhussein, ―Automatic Fruit Classification Using Deep Learning for Industrial Applications,‖ IEEE Access, vol. 8, pp. 140548–140556, Jul. 2020. doi: 10.1109/ACCESS.2020.3013612

work page arXiv 2020
[4]

D. C. Ramcharan, P. McCloskey, A. Baranowski, et al., ―A mobile -based deep learning model for plant disease diagnosis,‖ Frontiers in Plant Science , vol. 10, Mar. 2019, Art. 272. doi: 10.3389/fpls.2019.00272

work page arXiv 2019
[5]

Genctur k, S

A. Genctur k, S. Arslan, and M. Alkan, ―Maturity classification of dragon fruits using deep learning methods,‖ Sustainability , vol. 16, no. 3, p. 1342, 2024. doi: 10.3390/su16031342

work page doi:10.3390/su16031342 2024
[6]

J. Chen, L. Li, and X. Zhao, ―Dragon fruit maturity detection based on YOLOv5,‖ Agriculture , vol. 12, no. 11, p. 1794, Nov. 2022. doi: 10.3390/agriculture12111794

doi:10.3390/agriculture12111794 2022
[7]

X. Liu, Q. Wang, and Y. Wang, ―A Lightweight and Rapid Dragon Fruit Detection Method for Harvesting Robots,‖ Agriculture , vol. 15, no. 11, p. 1120, Nov. 2025. doi: 10.3 390/agriculture15111120

work page 2025
[8]

Howard et al

A. Howard et al. , ―MobileNets: Efficient convolutional neural networks for mobile vision applications,‖ arXiv preprint arXiv:1704.04861, Apr. 2017. [Online]. Available: https://arxiv.org/abs/1704.04861

arXiv 2017

Show all 18 references

[9]

Tan and Q

M. Tan and Q. Le, ―EfficientNet: Rethinking model scaling for convolutional neural networks,‖ in Proc. Int. Conf. Mach. Learn. (ICML) , 2019, pp. 6105–6114

2019
[10]

Islam, S

M. Islam, S. Bhuiyan, and S. Hasan, ―LeafNet: Channel Att ention-Driven Optimized CNN for Efficient Early Detection of Plant Diseases,‖ Agriculture , vol. 15, no. 2, p. 127, Jan. 2025. doi: 10.3390/agriculture15020127

2025 doi
[11]

Tonmoy, T

M. Tonmoy, T. Khan, and S. Barman, ―MobilePlantViT: A Mobile -friendly Hybrid Vision Transfo rmer for Generalized Plant Disease Image Classification,‖ arXiv preprint arXiv:2503.16628, Mar. 2025. [Online]. Available: https://arxiv.org/abs/2503.16628

2025 arXiv
[12]

Kumar and S

R. Kumar and S. Gupta, ―MobileViTV2: An Enhanced Lightweight Vision Transformer for Plant Disease Detection,‖ arXiv preprint arXiv:2408.01752, Aug. 2024. [Online]. Available: https://arxiv.org/abs/2408.01752

2024 arXiv
[13]

Singh, R

S. Singh, R. Kumar, and A. Choudhary, ―Tomato leaf disease detection using MobileNetV2,‖ International Journal of Computer Applications , vol. 175, no. 30, pp. 1–7, 2020. doi: 10.5120/ijca2020920131

2020 doi
[14]

Islam, M

M. Islam, M. Alam, and S. Rahman, ―Rice leaf disease classification usin g improved MobileNetV2,‖ Electronics , vol. 11, no. 6, p. 972, Mar. 2022. doi: 10.3390/electronics11060972

2022 doi
[15]

Y. Li, L. Zhang, and H. Huang, ―Lightweight Coffee Fruit Ripeness Detection Model Based on YOLOv5,‖ Frontiers in Plant Science , vol. 14, p. 1484784, Jan. 2025. doi: 10.3389/fpls.2025.1484784

2025
[16]

M. H. Saleem, J. Potgieter, and K. M. Arif, ―A plant disease detection mobile application using deep learning,‖ Electronics , vol. 10, no. 12, p. 1424, Jun. 2021. doi: 10.3390/electronics10121424

2021 doi
[17]

Z. Li, W. Zhang, and M. Yang, ―Embedded YOLOv5 for orchard fruit detection on low -power devices,‖ Sensors, vol. 21, no. 18, p. 6234, Sep. 2021. doi: 10.3390/s21186234

2021 doi
[18]

Kamilaris, A

H. Kamilaris, A. Kartakoullis, and F. Prenafeta -Boldú, ―A review on the practice of b ig data analysis in agriculture,‖ Computers and Electronics in Agriculture , vol. 143, pp. 23 –37, Dec. 2017. doi: 10.1016/j.compag.2017.09.037

2017 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Zhang, Y

W. Zhang, Y. Cao, H. Zhang, and Z. Li, ―Light -weight Fruit -Detection Alg orithm for Edge Computing Applications,‖ Frontiers in Plant Science , vol. 12, Oct. 2021, Art. 740936. doi: 10.3389/fpls.2021.740936

work page arXiv 2021

[2] [2]

Kamilaris and F

H. Kamilaris and F. X. Prenafeta -Boldú, ―Deep learning in agriculture: A survey,‖ Computers and Electronics in Agricult ure, vol. 147, pp. 70–90, Apr. 2018. doi: 10.1016/j.compag.2018.02.016

doi:10.1016/j.compag.2018.02.016 2018

[3] [3]

M. A. Khan, M. Y. Javed, and R. Alhussein, ―Automatic Fruit Classification Using Deep Learning for Industrial Applications,‖ IEEE Access, vol. 8, pp. 140548–140556, Jul. 2020. doi: 10.1109/ACCESS.2020.3013612

work page arXiv 2020

[4] [4]

D. C. Ramcharan, P. McCloskey, A. Baranowski, et al., ―A mobile -based deep learning model for plant disease diagnosis,‖ Frontiers in Plant Science , vol. 10, Mar. 2019, Art. 272. doi: 10.3389/fpls.2019.00272

work page arXiv 2019

[5] [5]

Genctur k, S

A. Genctur k, S. Arslan, and M. Alkan, ―Maturity classification of dragon fruits using deep learning methods,‖ Sustainability , vol. 16, no. 3, p. 1342, 2024. doi: 10.3390/su16031342

work page doi:10.3390/su16031342 2024

[6] [6]

J. Chen, L. Li, and X. Zhao, ―Dragon fruit maturity detection based on YOLOv5,‖ Agriculture , vol. 12, no. 11, p. 1794, Nov. 2022. doi: 10.3390/agriculture12111794

doi:10.3390/agriculture12111794 2022

[7] [7]

X. Liu, Q. Wang, and Y. Wang, ―A Lightweight and Rapid Dragon Fruit Detection Method for Harvesting Robots,‖ Agriculture , vol. 15, no. 11, p. 1120, Nov. 2025. doi: 10.3 390/agriculture15111120

work page 2025

[8] [8]

Howard et al

A. Howard et al. , ―MobileNets: Efficient convolutional neural networks for mobile vision applications,‖ arXiv preprint arXiv:1704.04861, Apr. 2017. [Online]. Available: https://arxiv.org/abs/1704.04861

arXiv 2017