Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

It is recommended to point out the core contribution of the model improvement in the abstract: "The improved model we proposed has improved the accuracy of infrared image detection by 2%, reduced parameters by 75%, and increased processing speed by 12.8%." In addition, the following suggestions are also put forward for the author's reference.

1. The literature review mainly focuses on comparing 1-stage and 2-stage detectors. There is insufficient analysis of the limitations of existing infrared image detection technology, and the necessity of research has not been highlighted. It is recommended to add relevant literature from electronics journals. Or articles recently published by mdpi. For example:
Aircraft Skin Damage Visual Testing System Using Lightweight Devices with YOLO: An Automated Real-Time Material Evaluation System
(The above is for reference only)
2. Conclusion Although the potential applications of the model are mentioned, no specific suggestions are given for future research directions, such as how to further improve the accuracy and efficiency of infrared detection. It is recommended to add specific future research directions in the conclusion; for example, we will study how to combine multimodal data to further improve the application scope of the model.

3. The description of parameter settings (such as hyperparameters of the GAM structure), data preprocessing and training strategies in the experiment is insufficient, which may affect the reproducibility of the experiment. It is recommended to add details of the experimental design, such as specific improvements to GAM, how to adjust the learning rate, and the optimizer selection during training.
4. Explain the data below each chart: "Figure 5 shows that the detection accuracy of the improved GAM structure in the campus scene is improved to 96%."

Author Response

We would like to thank the reviewers for their valuable comments and suggestions, which have helped us improve the quality of our manuscript.

Comment 1. The literature review mainly focuses on comparing 1-stage and 2-stage detectors. There is insufficient analysis of the limitations of existing infrared image detection technology, and the necessity of research has not been highlighted. It is recommended to add relevant literature from electronics journals. Or articles recently published by mdpi. For example:

Aircraft Skin Damage Visual Testing System Using Lightweight Devices with YOLO: An Automated Real-Time Material Evaluation System

(The above is for reference only)

Response 1) We added additional research on object detection in IR images and analyzed the drawbacks of those studies in “Introduction” section on page 1 from lines 40 to 51 as follows:

“However, IR images typically have lower resolution, noise issues, and smaller object sizes. Therefore, research and development of specialized object detection models for infrared image data is necessary. Current research on IR object detection focuses on two main approaches. The first approach involves noise filtering methods to improve image quality [6][7][8]. However, these techniques only address noise reduction without fully resolving resolution and object size issues. The second approach uses super-resolution techniques to enhance image resolution [6][9]. However, implementing both super-resolution and object detection simultaneously increases computational complexity, making it challenging for real-time processing, especially on devices with limited computing power. These limitations create a need for developing specialized object detection models for infrared images that can process efficiently in low-light conditions.”

 

Comment 2. Conclusion Although the potential applications of the model are mentioned, no specific suggestions are given for future research directions, such as how to further improve the accuracy and efficiency of infrared detection. It is recommended to add specific future research directions in the conclusion; for example, we will study how to combine multimodal data to further improve the application scope of the model.

Response 2) In Section 5 Conclusions of the paper, we added future research directions to lines 389 to 392 on page 12.

Specifically, as follows:” However, this paper only focuses on research in laboratory environments, schools, and human class detection. To make it applicable in real-world scenarios, the research scope needs to be expanded to other environments beyond schools, such as residential areas, traffic roads, etc., with a wider variety of classes.”

Comment 3. The description of parameter settings (such as hyperparameters of the GAM structure), data preprocessing and training strategies in the experiment is insufficient, which may affect the reproducibility of the experiment. It is recommended to add details of the experimental design, such as specific improvements to GAM, how to adjust the learning rate, and the optimizer selection during training.

Response 3) To address the concern about insufficient parameter settings and experimental details, we added comprehensive information in Section 4 (Experiments and Results) to lines 266 to 272 on page 8.

Specifically, as follows: “In this paper, the proposed system was trained under the following learning parameter conditions: Learning Rate, Decay, Channels, Epochs, Batch Size and Optimizer. The specific hyperparameter settings for the experiment are shown in Table 3.”

 

Table 3. Hyperparameter settings

Hyperparameters	Value
Input	640 x 480
Learning Rate	0.001
Decay	0.0005
Channels	1
Epochs	100
Batch size	4
Optimizer	Adam

 

Comment 4. Explain the data below each chart: "Figure 5 shows that the detection accuracy of the improved GAM structure in the campus scene is improved to 96%."

Response 4) Figure 5 in my article is "The Lightweight Architecture of YOLOv8." If you need an explanation of Figure 5, we have provided a detailed explanation from line 162 to 170 on page 5 as follows:

“YOLO models, particularly YOLOv8, achieve high precision with fast processing speeds. However, infrared images generally have the following characteristics: low resolution and smaller objects. In this paper, we propose model performance optimization for application on memory-constrained edge devices. We propose removing the P5 layer, which is one of three layers responsible for detecting large objects, from YOLOv8's head section. We also removed layers related to large object detection from the backbone and neck portions. Since P5 was originally designed for large objects and high resolution, it is unnecessary for object detection in infrared images.[6][10]”

 

Also, we have added a more detailed explanation of the results in Figures 10 and 11 (from line 323 to 354, page 11,12), illustrating that the proposed model detects objects more accurately compared to existing models, which often fail to detect them correctly.

Additionally, we have added the following explanation from line 355 to 368 on page 12.

“From the results shown in Figures 10 and 11, we evaluate that both YOLOv8s+P5 removal+ I_GAM and YOLOv8s+P5 removal+ I_GAM+ DW models have similar confidence scores, with no overlapping detections or missed objects. Another notable point is that the first three models: 1. Yolov8s 2. Yolov8s+P5 removal 3. Yolov8s+P5 removal +GAM tend to miss objects or excessively detect incorrect objects. Over-detection typically appears in two forms: multiple bounding boxes overlapping on a single object and false positives - marking areas without objects as if they contained objects. This is clearly demonstrated in Figures 10 and 11. In contrast, our proposed model has very balanced Precision (96%) and mAP (96.9%), showing that this model detects objects more accurately and minimizes the over-detection problem. Furthermore, when considering the processing speed from the results in Table 4, we can see that our proposed model not only achieves high accuracy and good mAP@50 with very few parameters but also maintains good confidence scores in the result images and fast processing speed, enabling accurate and efficient object detection. “

Reviewer 2 Report

Comments and Suggestions for Authors

I would like to thank the authors for submitting their manuscript entitled “Lightweight YOLOv8 Network for Thermal Image Object Detection”, which presents a lightweight object detection model based on YOLOv8, optimized for thermal imaging. Accuracy and speed are improved by removing unnecessary layers, integrating an improved GAM attention mechanism and using depthwise convolution. The resulting model achieves high accuracy (96%) and fast processing speed (55.86 fps) with significant parameter reduction, making it suitable for devices with limited computational resources.

The authors’ idea is clearly explained, and experimental results are promising. However, there are several aspects that would benefit from further clarification and refinement. Addressing the following comments will strengthen the manuscript and enhance its overall quality.

Comments and suggestions to authors

1. The references may be somewhat limited and are also a little out of date. The most recent is from 2023, followed by another from 2022, 2021.... In this field there are many related and recent works. Also, the references do not follow the proper format, lacking the DOI, URL, etc.

2. The “Author Contributions” section is missing, as well as the “Conflicts of Interest” statement. Neither is the “Data Availability Statement”, where it would be interesting to include the instructions to access the dataset they have built for their work.

3. Section 3. Proposed Method is rather long and contains too much information. Proposed Method, 4. Results and 5. Discussion.

4. Have you considered sharing the code and dataset used in the research for better reproducibility of the experiment?

5. It would be appropriate to include a detailed description of the image preprocessing, including the parameters of the convolutional filters or other enhancement techniques.

6. It is suggested to provide more details about the hyperparameters used during model training. Also, it would be useful to mention the type of optimizer, the learning rate, the batch size and the number of epochs used, as well as any other regularization techniques used.

7. In Table 2 there could be an error in the temperature range of the chamber used, since the text “~+500” is shown, which seems incomplete. Could you please indicate the range with its minimum-maximum values? If not, please provide some clarification.

8. Several model evaluation metrics are mentioned such as accuracy, FPS, or GFLOPs, could you please explain in more detail how they were calculated and what they mean? Also include information on the use of precision to evaluate model accuracy.

9. Regarding the good results obtained, it would be beneficial to evaluate the performance of the model with other data sets and other classes of objects, to validate its generalization. In fact, the evaluation of the model could be extended to different environments, in addition to the indoor environments (corridors, laboratories, classrooms and school areas) used in the study, to verify its robustness. Have they evaluated how the model would behave with other types of objects and scenarios (currently focused on Person detection)?

10. In the abstract they state that “The proposed model achieved high accuracy of 96% (a 2% improvement) and fast processing speed of 55.86fps (12.8% increase), while reducing parameters by approximately 75% compared to the original YOLOv8s”. This data sounds good, but then it does not show up well in the results shown in Figures 10 and 11. Could you please extend the discussion of the experimental results by elaborating on the reasons why your model performs better than the reference models in each of the scenarios?

11. In the new Discussion section, it would also be appropriate to clearly identify the limitations of the proposed model, such as its dependence on a specific data set, and to suggest possible future research directions beyond the examples cited in the text. It would also be good to analyze the impact of your contribution, highlighting how it differs from previous work and how it can influence practical applications. It is important to argue more explicitly the novelty of the proposal.

Author Response

We would like to thank the reviewers for their valuable comments and suggestions, which have helped us improve the quality of our manuscript.

Comment 1. The references may be somewhat limited and are also a little out of date. The most recent is from 2023, followed by another from 2022, 2021.... In this field there are many related and recent works. Also, the references do not follow the proper format, lacking the DOI, URL, etc.

Response 1) We added recent studies and also revised the citation format, including adding DOI links as follows:

5. Tausif Diwan, G. Anirudh, Jitendra V. Tembhurne. Object detection using YOLO: challenges, architectural successors, datasets and applications. SpringerNature Link. 2023, Volume82, pages 9243-9275. https://doi.org/10.1007/s11042-022-13644-y
6. Hyun-Seok Yoon, Eung-Tae Kim. YOLOv5-based Lightweight Methodology for Real-time Thermal Image Object Detection for Edge Devices. JOURNAL OF BROADCAST ENGINEERING. 2024, Vol. 29, No. 5, pp.703-712. https://doi.org/10.5909/JBE.2024.29.5.703
7. Xu Chen, Lei Liu, Jingzhi Zhang, Wenbo Shao. Infrared image denoising based on the variance-stabilizing transform and the dual-domain filter. In ScienceDirect. June 2021, Volume 113, 103012. https://doi.org/10.1016/j.dsp.2021.103012
8. Haitao Yang, Yaonan Tong, Zhiqi Cao and Zhenguang Chen. Infrared image enhancement algorithm based on improved wavelet threshold function and weighted guided filtering.in The 2nd International Conference on Electronic Information and Communication Engineering (EICE 2023) 13/01/2022 - 15/01/2022 Guangzhou, China. 2023. Microsoft Word - 223010517153931879

• Ning Zhang, Junnan Tan, Kaichun Yan and Sang Feng. Lightweight UAV Landing Model Based on Visual Positioning. In MDPI. Sensors 2025, 25, 884. https://doi.org/ 10.3390/s25030884
• Google Developers, Available online: Classification: Accuracy, recall, precision, and related metrics  |  Machine Learning  |  Google for Developers (accessed on February 4, 2025).

 

 

Comment 2. The “Author Contributions” section is missing, as well as the “Conflicts of Interest” statement. Neither is the “Data Availability Statement”, where it would be interesting to include the instructions to access the dataset they have built for their work.

Response 2) We added the following in Section 5: Conclusions, from lines 394 to 409 on pages 13 as follows:

“Author Contributions: Conceptualization, Eung-Tae Kim and Huyen Trang Dinh;  methodology, Eung-Tae Kim and Huyen Trang Dinh; software, Huyen Trang Dinh; validation, Huyen Trang Dinh; formal analysis, Huyen Trang Dinh; investigation, Huyen Trang Dinh; resources, Eung-Tae Kim; data curation, Huyen Trang Dinh; writing—original draft preparation, Huyen Trang Dinh; writing—review and editing, Eung-Tae Kim; visualization, Huyen Trang Dinh; supervision, Eung-Tae Kim; project administration, Eung-Tae Kim; funding acquisition, Eung-Tae Kim. All authors have read and agreed to the published version of the manuscript.

Funding: This work is supported by project for Industry-University-Research Institute platform cooperation R&D funded by the Ministry of SMEs and Startups in 2022.(S3312736)

Data Availability Statement: The dataset used in this study was developed through a government-funded research project in collaboration with industry partners. The availability of this data will be reviewed at a later date considering proprietary and confidentiality requirements.

Conflicts of Interest: The authors declare no conflict of interest.”

 

Comment 3. Section 3. Proposed Method is rather long and contains too much information. Proposed Method, 4. Results and 5. Discussion.

Response 3) We have reorganized the structure of the paper with the following sections:

1. Introduction
2. Related Work in Object Detection
3. Proposed System
4. Experiments and Results
5. Conclusions

Comment 4. Have you considered sharing the code and dataset used in the research for better reproducibility of the experiment?

Response 4) We added the following in Section 5 Conclusions, from lines 405 to 408 on pages 13 as follows:

 “Data Availability Statement: The dataset used in this study was developed through a government-funded research project in collaboration with industry partners. This data will be available based on request considering proprietary and confidentiality requirements.”

 

Comment 5. It would be appropriate to include a detailed description of the image preprocessing, including the parameters of the convolutional filters or other enhancement techniques.

Response 5) The dataset images used in the research were directly captured from the QuantumRed VR infrared camera with a resolution of 640x480. They were then labeled through Roboflow and input into the model for training, without going through any preprocessing steps such as cropping, resizing, color normalization, or noise cleaning.

Comment 6. It is suggested to provide more details about the hyperparameters used during model training. Also, it would be useful to mention the type of optimizer, the learning rate, the batch size and the number of epochs used, as well as any other regularization techniques used.

Response 6) We added comprehensive information in Section 4 (Experiments and Results). from lines 266 to 272 on page 8.

Specifically, as follows: “In this paper, the proposed system was trained under the following learning parameter conditions: Learning Rate, Decay, Channels, Epochs, Batch Size and Optimizer. The specific hyperparameter settings for the experiment are shown in Table 3

 

Table 3. Hyperparameter settings

Hyperparameters	Value
Input	640 x 480
Learning Rate	0.001
Decay	0.0005
Channels	1
Epochs	100
Batch size	4
Optimizer	Adam

 

 

Comment 7. In Table 2 there could be an error in the temperature range of the chamber used, since the text “~+500” is shown, which seems incomplete. Could you please indicate the range with its minimum-maximum values? If not, please provide some clarification.

Response 7) We added the minimum-maximum values in Table 2 on page 8.

 Table 2. QuantumRed VR Specifications:

Temperature Range

-40 to ~ 500˚C

 

Comment 8. Several model evaluation metrics are mentioned such as accuracy, FPS, or GFLOPs, could you please explain in more detail how they were calculated and what they mean? Also include information on the use of precision to evaluate model accuracy.

Response 8) We added detailed explanations of the model evaluation metrics mentioned, including: the number of parameters, GFLOPs, FPS, precision and mAP.  We also cited additional references regarding the meaning of model evaluation metrics and calculation methods as follows in Section 4: Experiment and Results, from lines 273 to 299 on page 9:

“We evaluate the performance of the models based on the following evaluation metrics: number of parameters, computational load (GFLOPs, Giga Floating Point Operations Per Second), FPS (Frames Per Second), precision and  (Mean Average Precision) [28][29].

The number of parameters used in deep learning model affects complexity and deploy ability on resource-constrained devices.

GFLOPs is calculated based on the number of floating-point operations required for matrix multiplications, convolutions, and other computations. It reflects computational workload, helping evaluate energy efficiency and suitability for real-time applications.

FPS refers to the number of frames processed per second. In deep learning and computer vision, FPS is used to measure the processing speed of models, such as object detection systems, indicating how many frames a model can analyze per second.

mAP (Mean Average Precision) is a widely used evaluation metric for object detection tasks, calculated as:

  mAP = 1/n ∑ AP_i (4)

where  represents the average precision (the area under the Precision-Recall curve) for the i-th category and n is the number of categories. In this paper, we use mAP@0.5 to evaluate our model, which represents the average of APs across all categories when the Intersection over Union (IoU) threshold is set to 0.5 [28].

Precision represents the reliability of positive predictions, particularly important in applications demanding high precision [29][30]. It is calculated according to Equation (5).         

  P = TP / (TP + FP) (5)

where TP(True Positives) is correctly predicted positive instances and FP(False Positives) is incorrectly predicted positive instances. In deep learning, precision is a metric used to evaluate the performance of a classification model, especially in cases of imbalanced datasets, it measures the proportion of correctly predicted positive instances among all instances predicted as positive. ”

 

Comment 9. Regarding the good results obtained, it would be beneficial to evaluate the performance of the model with other data sets and other classes of objects, to validate its generalization. In fact, the evaluation of the model could be extended to different environments, in addition to the indoor environments (corridors, laboratories, classrooms and school areas) used in the study, to verify its robustness. Have they evaluated how the model would behave with other types of objects and scenarios (currently focused on Person detection)?

Response 9) We also observed that the research has only been limited to a school environment and focused on a single class, which is 'person.' We added the following in Section 5: Conclusions of the paper, from lines 389 to 392 on page 12:

 “However, this paper only focuses on research in laboratory environments, schools, and human class detection. To make it applicable in real-world scenarios, the research scope needs to be expanded to other environments beyond schools, such as residential areas, traffic roads, etc., with a wider variety of classes.”

Comment 10. In the abstract they state that “The proposed model achieved high accuracy of 96% (a 2% improvement) and fast processing speed of 55.86fps (12.8% increase), while reducing parameters by approximately 75% compared to the original YOLOv8s”. This data sounds good, but then it does not show up well in the results shown in Figures 10 and 11. Could you please extend the discussion of the experimental results by elaborating on the reasons why your model performs better than the reference models in each of the scenarios?

Response 10) We added a more detailed explanation of the results in Figures 10 and 11 (from line 323 to 354, page 11,12), illustrating that the proposed model detects objects more accurately compared to existing models, which often fail to detect them correctly.

Additionally, I have added the following explanation from line 355 to 368 on page 12.

“From the results shown in Figures 10 and 11, we evaluate that both YOLOv8s+P5 removal+ I_GAM and YOLOv8s+P5 removal+ I_GAM+ DW models have similar confidence scores, with no overlapping detections or missed objects. Another notable point is that the first three models: 1. Yolov8s 2. Yolov8s+P5 removal 3. Yolov8s+P5 removal +GAM tend to miss objects or excessively detect incorrect objects. Over-detection typically appears in two forms: multiple bounding boxes overlapping on a single object and false positives - marking areas without objects as if they contained objects. This is clearly demonstrated in Figures 10 and 11. In contrast, our proposed model has very balanced Precision (96%) and mAP (96.9%), showing that this model detects objects more accurately and minimizes the over-detection problem. Furthermore, when considering the processing speed from the results in Table 4, we can see that our proposed model not only achieves high accuracy and good mAP@50 with very few parameters but also maintains good confidence scores in the result images and fast processing speed, enabling accurate and efficient object detection. “

 

Comment 11. In the new Discussion section, it would also be appropriate to clearly identify the limitations of the proposed model, such as its dependence on a specific data set, and to suggest possible future research directions beyond the examples cited in the text. It would also be good to analyze the impact of your contribution, highlighting how it differs from previous work and how it can influence practical applications. It is important to argue more explicitly the novelty of the proposal.

Response 11) In Section 5: Conclusions, we pointed out the difference between this research and previous studies: earlier studies focused on RGB images, while RGB images are quite sensitive under different lighting conditions. Therefore, researching models that use thermal images for recognition is necessary. Additionally, for real-time recognition applications on devices with limited GPU capabilities, models need to be both accurate and lightweight to ensure speed. This paper proposes a model that meets both of these requirements. We also provide the limitations and future research directions, which are from lines 389 to 392 on page 12 as follows:

  “this paper only focuses on research in laboratory environments, schools, and human class detection. To make it applicable in real-world scenarios, the research scope needs to be expanded to other environments beyond schools, such as residential areas, traffic roads, etc., with a wider variety of classes.”

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Although the manuscript has been improved, there is no new section explaining the differences and innovations between the references and this manuscript. In addition, the literature citations do not follow the author guidelines. It is recommended to refer to: Aircraft Skin Damage Visual Testing System Using Lightweight Devices with YOLO: An Automated Real-Time Material Evaluation System. To meet the format requirements.

Author Response

We appreciate the Editor and the Reviewer for their detailed and constructive comments on the manuscript.

For the revision, we tried to clarify all the comments raised in the review. In the following, we describe how we have addressed the raised issues. The corresponding changes of the revised manuscript are highlighted in red in the final manuscript.
 Our responses to the comments are described in a point-by-point manner and attached in the file below.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors,

Thank you for your efforts to improve the manuscript and for incorporating all my recommendations. A substantial improvement of the work can be appreciated.

I recommend your acceptance.

Best regards,

Author Response

We appreciate the Editor and the Reviewer for their detailed and constructive comments on the manuscript.

Thank you once again for your acceptance.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have improved their manuscript according to the reviewers' comments. It can be accepted for publication in this version.