All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
The paper may be accepted.
[# PeerJ Staff Note - this decision was reviewed and approved by Shawn Gomez, a PeerJ Section Editor covering this Section #]
The authors have incorporated all my comments.
-
-
The authors have meticulously addressed the comments raised by the reviewer, and now the manuscript can be published.
-
-
The authors made the revisions.
Experimental design part revised.
The authors discuss the performance metrics deeply.
The scope of the paper is extended.
Validity of the results is checked.
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
Introduction
• Present a clear research gap: why do existing transfer-learning models underperform on fruit images? Merely citing accuracy numbers from small, dated datasets is insufficient.
• Condense general facts about supermarket bar-code systems; focus instead on the unique visual challenges (occlusion, lighting, similar colors) that motivate your study.
• End with a concise list of concrete contributions aligned with what is actually new (e.g., dual-dataset evaluation, systematic augmentation study, if added).
Literature Review
• Organize chronologically or thematically rather than listing 25 papers in sequence.
• Add recent fruit-classification works that use Vision Transformers, EfficientNet-V2, or MobileViT (2024-25) and discuss why VGG-19 is still chosen despite higher parameter count.
• Include a comparison table (dataset, classes, split protocol, top-1 accuracy) to show where your numbers sit relative to the state of the art.
• Several citations are duplicated: tidy references.
Methodology / Approach
• Provide a schematic or table of the exact layer freezing strategy (you state “first ten conv layers were frozen” for Dataset 2, but the freeze policy for Dataset 1 is unclear).
• Detail augmentation settings (rotation degrees, flip probability, color-jitter ranges) so others can replicate.
• Explain why an adaptive SGD schedule was chosen over commonly used Adam/AdamW; justify the decay formula hyperparameters.
• Discuss class imbalance (e.g., Cucumber has only 200 images); did you use weighted loss or oversampling?
• Release code or at least a GitHub link with training scripts and the exact train/val/test file lists.
Experimental Results
• Report 5-fold cross-validation or, at a minimum, repeat each experiment three times with different seeds and show mean ± SD.
• Add precision, recall, F1, and per-class accuracies; high macro-average numbers can hide poor minority-class performance.
• The confusion matrices are useful; add numeric totals or percentage labels for interpretability.
• Discuss the 22 misclassifications in Dataset 1: what visual traits led to Avocado↔Mango errors? Provide sample images.
Experimental Evaluation
• Provide inference latency on a resource-constrained device (e.g., Jetson Nano or Raspberry Pi) to support “real-time” claims; Kaggle T4 GPU speed is not representative.
• Compare against lighter models (MobileNetV3-Small, EfficientNet-Lite) fine-tuned under the same protocol to justify VGG-19’s extra parameters.
• Include a small ablation study: no augmentation, different learning-rate decay, smaller image size (64 × 64 px) to quantify each design decision.
• Perform a train-from-scratch baseline to show the benefit of transfer learning.
Conclusion
• Update conclusions: accuracy is impressive under the chosen split, but generalization to other fruit datasets, lighting, or devices is untested.
• Explicitly list limitations: single data source, image-level labels, no real-world deployment test, and possible data leakage.
• Outline specific future work: cross-orchard dataset, domain adaptation, lightweight on-device inference, public code release.
The manuscript fine-tunes VGG-19 on two publicly available Kaggle “Fruit-360” subsets and reports headline accuracies of 99.04 % and 97.58 %. While automated fruit recognition is useful for retail and dietary-tracking applications, the study’s contribution is modest, and several weaknesses limit its credibility:
• Using a pre-trained VGG-19 with standard transfer-learning is well-trodden ground; no architectural change or new learning trick is introduced.
• Both datasets come from the same online source, so the test images are not truly “unseen.” There is only one 75 / 10 / 15 split, no cross-validation, no subject-level partitioning, and no external benchmark; yet near-perfect metrics are claimed.
• The paper does not state whether images from the same fruit instance (or highly similar augmentations) appear in both train and test sets, which could hugely inflate accuracy.
• Key hyperparameters (optimizer settings beyond the adaptive-SGD schedule, augmentation probabilities, early stopping, class-balance handling) are missing. Code and data-split files are not released, hindering reproducibility.
• The manuscript repeatedly says the model “outperforms all existing methods,” yet compares only with three older baselines and omits recent (2024-25) transformer or lightweight CNN approaches.
The manuscript presents a novel and engaging subject with a well-structured framework and clear presentation. However, several revisions are necessary before it can be considered for publication.
1- Lines 34-40: Please provide a reference for this paragraph.
2- Lines 41-43: Please provide a reference for this statement.
3- lines 209 to 210: The reason for using dropout layers should be supported by a reference.
4- Line 262: The authors should justify the decision to split the dataset into training and test subsets without incorporating a validation subset. While maintaining the current structure, the authors are encouraged to discuss the technical reasoning behind this choice, as it is seldom addressed in the literature.
-
-
1. Grammar of the paper should be checked.
2. The related work part should be extended.
3. In the abstract, there should be a sentence for novelty.
4. Authors should read papers related to smart agricultural applications.
5. There should be contributions and research objectives in the introduction.
1. Is this method realistic to use in real-time applications?
2. Please check all figures, some of which are so unprofessional.
3. Figure 4 shows samples of the fruit pixel quality not well.
4. Why do authors choose these datasets for applications?
5. VGG16 model, there should be a citation for this if the authors did not propose this architecture.
6. Are the metrics enough to compare?
1. The conclusion part should be revisited with future directions.
2. Authors should also give other metrics to compare with recall, etc.
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.