Performance Comparison of Cherry Tomato Ripeness Detection Using Multiple YOLO Models

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Major Revision Suggestions:  
The paper requires major revisions to improve its scientific rigor, novelty, and practical relevance. Key areas for improvement include the following:  

1. Novelty and Contribution: The combination of YOLOv8 and ResNet50 lacks significant novelty as the use of ResNet50 as a backbone has already been extensively explored in the fields of computer vision and agricultural automation. Comparative experiments with other popular backbone architectures, such as MobileNet or EfficientNet, should be conducted to justify the selection of ResNet50. Additionally, the descriptions of YOLO models are overly repetitive and should be streamlined, focusing on the unique contributions and specific innovations of the study.

2. Dataset and Realism: The dataset used in this study is relatively small, with only 300 images expanded to 742 through data augmentation. This limited scale is insufficient to represent the complexity and diversity of real-world agricultural scenarios. It is recommended to expand the dataset and include experiments to verify whether the augmented data effectively simulate environmental variations encountered in real-world settings. Furthermore, detection results under challenging conditions such as varying lighting, occlusions, and complex backgrounds should be included to better assess the robustness of the proposed model.

3. Experimental Design and Analysis: The performance analysis in the paper lacks depth. More quantitative and qualitative insights into the observed performance differences across models are needed, particularly regarding why ResNet50 improves YOLOv8’s performance in terms of feature extraction and model robustness. Detailed evaluations of failure cases, such as false positives and false negatives, should also be provided to better illustrate the model’s limitations and areas for improvement.

4. Experimental Settings and Reproducibility: The paper outlines the experimental setup but lacks sufficient details on hyperparameter tuning, such as learning rate schedules, optimizer selection, and data augmentation strategies, which are crucial for reproducibility. Evaluating the model’s performance in dynamic or real-world agricultural environments, including factors such as target movement and fluctuating lighting conditions, would also strengthen the study’s applicability.

5. Practical Applications: While the paper discusses future applications in harvesting robots, it lacks validations in real-world deployment scenarios. Real-time tests in practical agricultural environments, simulating real-world factors like occlusions and background clutter, should be incorporated to substantiate the claims of applicability. Furthermore, segmentation techniques to enhance ripeness detection granularity, as mentioned in the conclusions, should be explored and tested for feasibility in the current context.

Author Response

첨부파일을 참고해 주시고, 회원 여러분께 진심으로 감사드립니다.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

After review, the paper has the following deficiencies:

1. The abstract is not written in a standardized manner and lacks result data. It is recommended to revise it according to the purpose (background), methods, results, and conclusions;

2. The paper did not test the model in a real environment. In actual applications (lighting conditions, background interference) are also influencing factors, and its performance in actual applications is still unknown;

3. In “2. Related Research”, the cited literature is expressed as “in [11], in [12], in [13]”, etc. It is recommended to change it to “Khan et al. [11], Vo et al. [12], Li. et al. [13]”, etc.

4. 4.2 proposed the application of “loss function” to optimize the performance of the model. Only the concept of the function was given. It is recommended to supplement the specific model and related parameters of the function;

5. There are too few key model formulas given in the article;

6. The paper compares multiple YOLO models, but the experimental results show that the accuracy of YOLO V8s (ResNet50) is only 0.668, which is relatively low;

7. Although the paper applies data enhancement techniques (such as saturation adjustment, brightness change, exposure modification and image flipping, etc.) to reflect weather fluctuations and environmental conditions in agricultural environments, these methods are not sufficient to cope with the diversity and complexity of actual agricultural environments, and the dataset diversity is insufficient; it is recommended that the dataset include multiple conditions such as different time periods of the day, different seasons, different weather, different angles, and different sources (online collection, different cameras and their basic parameters).

8. The ratio of images using data augmentation technology to original images in the dataset is not specified. It is recommended to complete it.

9. The detection speed is not given, which does not reflect the work efficiency.

Comments on the Quality of English Language

Further polish the language expression.

Author Response

Please see the attachment, and we sincerely thank the reviewers for their valuable feedback.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

After the author's revision, some problems in the paper were corrected, the overall quality of the paper has been improved, and it has certain academic reference value.