Insulator Defect Detection Based on YOLOv8s-SwinT

Round 1

Reviewer 1 Report

I have the following concerns.

1. The use in ML of swin transformer block is known. Therefore, in section 2 you need to clarify your distinction.

2. In section 3, show which classes of defects you identify and classify.

3. As a rule, images of insulators are obtained during the survey of high-voltage transmission lines with the help of drones. It is known that real images are often blurred. How does your technology take this into account.

4. The cited metrics (1)-(4) are known in ML, so there is no point in citing them.

5. It is necessary to show on which databases the comparative results given in Table 3 were obtained.

6. State the limitations of your approach to detecting defects in insulators.

7. References must be significantly expanded and supplemented with articles for 2022-2024 to confirm the relevance of the problem statement.

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Insulator Defect Detection based on YOLOv8s-SwinT

This research paper proposes machine learning based detection of Insulator defects. The proposed method leverages combination of different components like Swin Transformer, YOLOv8s, BiFPN, C2fSTR module and weighted fusion of features, etc. A comparative analysis of proposed approach is also performed with other architectures like Faster R-CNN and different versions of YOLO versions. Accurate identification of faults in insulators can be beneficial to avoid accidents in an installed operating power system.

The paper is well-written, and content is well organized. However, following minor comments need to be addressed:

·         The details related to dataset are brief or missing like source of dataset, original resolution of images, etc.

o   If it is a publicly available dataset, then cite its link and if the images are collected by authors so it should be made public for experimentation and benchmarking purposes.

·         Overall, a nice comparative analysis is performed. Also add the training time, inference time and number of parameter flops in comparison table.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments for author File: Comments.pdf

The quality of the language is adequate but minor corrections should be made.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The authors proposed YOLOv8s based insulator defect detection by integration of SwinT attention module which is an interesting work. There are some technical comments that should be clearly addressed in the manuscript.

1). The used data for training and testing should be tabulated in terms of image size and image number.

2). What are the input image size?

3) The input size should be at least 1000 by 1000.

4). The integration of the recent YOLOV8 and Swin attention is good, but it is a damage detection not damage segmentation. Damage detection requires additional process that is quantification of damage. There are some recent state of the art methods with large input image size with at least 1000x 1000 and FPS is more than 50: SDDNet: Real-time crack segmentation; Efficient attention-based deep encoder and decoder for automatic crack segmentation; Attention-based generative adversarial network with internal damage segmentation using thermography.5

5) what are the new technical contribution of this paper compared to the previous suggested methods except the existing method application?

6) In this deep learning based damage/defect detection, there are tow original and key papers: Deep learning-based crack damage detection using convolutional neural networks; Autonomous structural visual inspection using region‐based deep learning for detecting multiple damage types. These should be discussed in the beginning of the introduction. Also there are extensive literature review in this deep learning based structural health monitoring: Deep learning-based structural health monitoring.

7). The mean intersection over union should be used as an evaluation metric.

8). For the insulator monitoring of transmission tower. UAV is essential. What kind of UAVs were used in data collection? There are some recent methods of autonomous UAVs with integration of advanced deep learning for damage segmentation with real time processing and mapping. Obstacle avoidance method for autonomous UAV for structural health monitoring; Real-time multiple damage mapping using autonomous UAV and deep faster region-based neural networks for GPS-denied structures. These essential works should be discussed to address recent advanced development in this specific topic.

Overall, it is interesting. But the novelty of the proposed method should be further clarified compared to the existing works suggested above, and overall literature reviews are very poor by neglecting many key papers and advanced works.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Overall, I am satisfied with the answers and additions to the article. 

Some English editing is required. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Even though the authors revised the paper but still the main issues are not resolved. The authors mentioned that quantification is not the goal of this paper and it is an emerging topic therefore they did not do that. However, it is not an emerging topic but it is mutured area of the research in this deep learning based methods. Therefore, the reviewer suggested some key papers but none of them were even discussed to address current issue.

Also the same way, for this transmission tower inspection, the use of UAV is essential. And some recent works with autonomous UAVs were suggested, but the authors never touch any recent development in this topic which makes the paper outdated because YOLO and attention have been used extensively already. I encourage the authors address these two main issues again.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 4 Report

all the comments were addressed.