Pipeline Landmark Classification of Miniature Pipeline Robot π-II Based on Residual Network ResNet18
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper presents a novel and intriguing procedure for identifying pipe landmarks in detail. The introduction comprehensively reviews the current state-of-the-art robotics and machine vision techniques used for pipeline landmark detection.
The proposed solution, a ResNet-18 network coupled with a cost-effective fisheye camera, offers a potentially simpler, cheaper, more robust, and reliable alternative compared to existing methods documented in the literature.
While the impact of "periodically occurring small sample sizes on training performance" and the influence of the "batch_size parameter" might not carry such a scientific novelty, the authors deserve credit for their detailed analysis of these factors.
The authors' conclusion regarding the need for a larger dataset is well-founded. Exploring synthetic data generation as future work could be particularly interesting.
The following questions emerged during the review process. While addressing them is not mandatory, they could be valuable for future research:
1. How resilient is the system to dirt or corrosion on the interior walls of the steel pipes?
2. The current approach necessitates a large dataset of images specific to each pipe, hindering its universal applicability. Addressing new scenarios requires data generation and retraining for each unique case.
3. Could the dataset have been expanded by incorporating noise and disturbances similar to [18] to increase the sample count?
The reviewer encountered a few potential translation errors and ambiguities regarding the use of definite and indefinite articles ("a/an," "the") while reading the manuscript. A comprehensive proofreading of the entire document is recommended. Specific instances of these issues are highlighted in the following "Minor Problems" section.
Major problems:
None
Minor observations regarding the manuscript:
1. While the content of Figure 10 appears logically consistent, for the sake of clarity and adherence to common conventions, the ReLU activation function at the top of the image might be more appropriately positioned after the "add" element.
2. It might be beneficial to consider cropping Figure 2 to remove any sections referencing SolidWorks. This would enhance the overall visual presentation.
3. The titles for sections 3.1.2, 3.1.3, 3.1.4, 3.2, and 4.4 appear at the end of their respective pages. While these are likely to be automatically corrected during final formatting, it's worth noting for a more polished presentation.
Comments on the Quality of English LanguageMinor problems (primarily grammatical errors):
1. In the sentence “… camera centrally located at the front end.” consider writing “camera located at the center of the front end.”
2. In the sentence “The paper provides a detailed analysis of the image characteristics … and creates a dataset …“, is the dataset created by the paper or the authors?
3. In the sentence “… the final test results indicate that this modified network has a high accuracy rate in classifying various pipeline landmarks, demonstrating the promising application prospects … ” The article “the” before “promising application prospects” refers to specific expectations. I think the use of the article "a" is more appropriate, as the sentence is about “promising application prospects” in general.
4. In the sentence “In a straight pipe, the near end close to the robot camera will have a very strong reflection”, what does “near end” mean? Did the authors mean “near wall”?
5. In the sentence “For T-shaped pipes, there are two modes”, did the authors mean “two cases”?
6. In the sentence “… π II-type pipeline robot, is developed …”, did the authors mean “was developed”?
7. In the sentence “… structures beyond 2 meters is photographed.”, did the authors mean “were photographed”?
8. In the sentence “show images taken at intermediate and farthest distances.”, I think, it should be „at the farthest”.
9. In the sentence „After preprocessing the captured images, use Python commands to …” I think, it should be „Python commands were used”, although the information itself (that Python commands were used) is not essential in this sentence.
10. In the sentence „... that connect both the original signal and the incremental signal.”, are the authors sure „incremental” is the right phrase? Shouldn’t it be „residual” or maybe „difference”?
11. In line 367 the word „where” is missing, I think.
12. Please reconsider rephrasing the sentence “In this paper, train for 100 epochs, and optimizes parameters continuously.”
13. In the sentence “Set the hyper parameter “batch_size” to 64 …” I think, it should say „were set to...”
14. The paragraph between lines 403-409 is hard to follow. Please consider rephrasing it.
15. In the sentence “For the pipeline landmark classification discussed in this paper, a better “batch_size””, what does “better” mean?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript proposed a ResNet18-based Pipeline Landmark Classifier for a wheel-legged mobile robot capable of moving inside the pipeline with a monocular fish-eye camera and LED lights. The dataset of 908 images containing pipeline landmarks is created and used to train the network.
The paper is well written. The methodology: datasets formation and training are well explained for the work that has been done. The theoretical contribution of this work is questionable. However, the work has practical significance for inspections of pipelines that are typically underground.
My concerns about the manuscript are as follows:
a) The authors should mention the contribution of the work in the introduction section. As per my understanding, the contribution is just creating the pipeline dataset.
b) The training curves questions about the selection of the ResNet18. It seems like the training data overfits the trained model. The authors should consider more extensive datasets to train the model or use other small & lightweight models like Yolo capable of running in real-time.
c) The evaluation of the trained model is very limited to just microaccuracy. More extensive evaluation is required using standard metrics for all classes. The paper should discuss the results for the cases such as false positives and true negatives.
d) It is better to summarize training and testing parameters in a table for better understanding.
e) The Python codes in the paper should be presented using a Monospaced font and may be outlined by box.
f) The instances like Figures 6(a) and (b) may be written as Figures 6(a) and 6(b) for consistancy.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper presents a pipeline robot, π-II, suitable for micro-pipeline inspection, and utilizes ResNet18 and other residual network models to learn from images of various pipeline landmarks captured by a fisheye camera. The proposed method is innovative, but I still have some concerns as follows:
1. The author's writing is not very standard, as it is rare to include actual code in formally published journal papers, which can severely impact the reader's experience. I suggest the author use pseudocode to express the algorithms used in the paper instead of actual code, though the author can modify this before the manuscript is published.
2. The author should include a framework diagram of the entire method, so readers can understand the work more intuitively.
3. There are some spelling and grammatical errors in the manuscript. The author should carefully review the manuscript and correct these errors, as this is important for improving the quality of the manuscript.
4. The author needs to improve the quality of some images, such as Figures 5, 6, 8, 11, 12, and 13.
5. The Conclusions section needs to be written in more detail.
6. The author should include more meaningful future work, which is important for the readers.
Comments on the Quality of English Language
1. There are some spelling and grammatical errors in the manuscript. The author should carefully review the manuscript and correct these errors, as this is important for improving the quality of the manuscript.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsAuthors introduced the robot system π-II for 120mm pipe inspection using a fisheye camera and DNN to classify 4 categories (landmarks)of inner pipelines features.
An interesting article is presented given the importance of these pipeline inspection robots. However, there are aspects that can be included to improve the readability of the paper and the appropriateness of the manuscript within the scope of this journal.
1. It is convenient to include a description of the robot's movement mechanism. How does it move from a straight-line pipe to a curve and how does it move when it reaches a T-shaped joint?
2. This type of robot uses an umbilical cord to transmit video and control data. The authors mention the importance of having a low runtime, but never mentioned about its value. How important this is depends on the computing capacity of the control computer. Could you elaborate on this matter, what are the computational resources available to the equipment and what are the runtime mentioned and which are necessary to effectively have real-time control?
3. It would be helpful if the authors made their image database public so that other researchers can compare similar developments in the field of feature recognition inside pipelines.
4. In section 11, the authors mention the 3-channel BRG camera. I am wondering if they meant the typical RGB signal. Please explain.
5. To evaluate the effectiveness of their method using Resnet, it is necessary to include the confusion matrix.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI thank the authors for the revised manuscript. However, the revision does not address my concern at least in an acceptable way. The paper's contribution is very minor i.e. just the creation of the pipeline dataset. It seems like the training data overfits the trained model as can be seen from the presented tables and curves. The training curves questions about the selection of the ResNet18. The authors should consider more extensive datasets to train the model or use other small & lightweight models that are suitable.
Author Response
Thank you for your suggestion. Creating a dataset is a very laborious task, and in the future work, we will expand such dataset.
Reviewer 4 Report
Comments and Suggestions for AuthorsThe authors appropriately include responses to the observations. Highlighting the main objective of this work, providing more references from their own work to describe the design of the robot and included the computing resources. Additionally, they provided a github address with their database and a confusion matrix for the results. The article is technically sound and can be published.
Author Response
Thank you for your affirmation.
