Machine Learning and Queuing Algorithm Integration for Real-Time Citrus Size Classification on an Industrial Sorting Machine

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents a practical and well-structured integration of a lightweight machine-learning pipeline and a microcontroller-based queuing algorithm for real-time lemon size classification on an industrial sorting machine. The authors clearly describe the motivation for replacing an obsolete vision system, and they demonstrate a full workflow from hardware integration to feature extraction, classifier training, and embedded deployment. The study addresses a relevant applied problem and shows promising results using a dataset collected directly from a functioning industrial line.

However, several aspects of the manuscript require clarification and strengthening before the work can be considered for publication.

1. Please specify the exact train/validation/test protocol, including fold count and whether images of the same fruit could appear in different splits. Clarify whether a fully held-out test set, untouched by GridSearchCV, was used.
2. Include additional evaluation metrics such as per-class precision, recall, F1 scores, macro/micro averages, balanced accuracy, and confidence intervals to provide a more complete view of classifier performance. 
3. Please, provide an ablation study to show the individual contribution of the peduncle/radius feature relative to geometric features alone, and compare the chosen CIELab a/b segmentation with at least one alternative (e.g., HSV). Also state the thresholding and contour filtering criteria used.
4. Describe the SMOTE procedure in detail, including the parameters used, confirmation that it was applied only to the training set, and per-class counts before and after oversampling. A small table summarizing performance changes for 0%, 20%, and 50% oversampling would be helpful.
5. Report quantitative real-time performance of the entire pipeline, including end-to-end latency.
6. Since testing was performed on an actual industrial sorter, please provide quantitative field results such as the number of fruits processed, mis-routing rate, downtime, operator interventions, and a confusion matrix from the live run. If possible, include a brief comparison to the legacy vision system
7. For the peduncle detection model, provide the exact training hyperparameters, dataset splits, and any augmentations used. Please also define what “vectorized” versus “non-vectorized” accuracy means in Table 2.
8. Indicate whether the dataset or portions of the code can be shared to support reproducibility. Additionally, Figures 6 and 8 would benefit from larger fonts and more descriptive captions to improve clarity. 

Author Response

Please read the attached file

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper proposes a real-time lemon size classification system based on machine learning and queueing algorithms, and integrates it into an industrial sorting machine, demonstrating practical utility. Here are some concerns that the authors should address during revision. 1. The authors mentioned in Lines 191-192, FOMO, MobileNetV2, YOLO-Pro were trained. However, no experimental results support the selection of the FOMO model. The authors should supplement comparative experimental results for the object detection task. 2. For balancing the dataset, SMOTE was used. What about its effect? I do not see any results in terms of this algorithm. 3. Still, in the Results section, the authors only presented evaluation indicators regarding detection precision. What about the efficiency, like FPS, system delay, memory usage? 4. For Section 2.5, as the authors have already obtained the features (e.g., contour area, contour perimeter, etc.), why using a classifier subsequently? One can easily grade the lemons directly with these features. 5. As the dataset was limited by 3127 samples, the stability and robustness of the proposed system can be hardly evaluated. Extensive experiments should be supplemented to see how this system work for a longer time. 6. The authors claimed this was a cost-effective machine vision system. However, no evidence supports this statement. 7. In the future work part, I doubt that the illumination condition would be an issue. Because within a sorting system, one can supplement passive lighting to maintain the stability of the system, rather than considering it as a variable.

Author Response

Please read the attached file

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I would like to thank the authors for their careful and thorough revision of the manuscript. The responses provided are clear and address the previously raised concerns in a satisfactory manner. The revised version shows improved clarity, methodological transparency, and overall presentation.

Reviewer 2 Report

Comments and Suggestions for Authors

I recommend accepting this manuscript.