A Railway Track Extraction Method Based on Improved DeepLabV3+

Round 1

Reviewer 1 Report

This proposed method for railway track extraction using an improved DeepLabV3+ model has several advancements, but upon closer examination, it still falls short in various aspects.

Firstly, while the introduction of MobileNetV3 as the backbone extraction network is touted as a "lightweight" enhancement, the claim is not fully supported with concrete evidence.

Secondly, adopting the CARAFE universal upsampling operator for the decoder part is an improvement. Still, the lack of comparative analysis with other upsampling techniques leaves doubt about its superiority and efficiency.

Furthermore, the morphological algorithms applied to optimize extraction results are claimed to address potential errors. However, the reviewer is concerned about the need for more detailed information about the specific algorithms used and how they can improve the accuracy of railway track extraction.

Although the experimental results demonstrate relatively high MIoU scores, Recall values, and overall accuracy on the dedicated railway track segmentation and DeepGlobe datasets, the reviewer questions the model's generalizability. The absence of a broader range of datasets and real-world examples raises doubts about the algorithm's effectiveness in diverse railway track extraction scenarios.

Figures 8 and 9 size and resolution should be improved.

Author Response

Dear Reviewer,

We hope this email finds you well. We sincerely appreciate your time and effort in reviewing our paper titled "A Railway Track Extraction Method Based on Improved DeepLabV3+." Your valuable feedback has played a crucial role in improving the quality and impact of our work.

We have carefully considered each of your comments and suggestions, and have made necessary revisions to address them. In this email, we provide detailed responses to each of your points:

 

1. Specific evidence supporting the performance improvement of MobileNetV3 is detailed in reference [19]. In this paper, we have made the following modifications: a summary of the reasons for the performance improvement of MobileNetV3 will be added after Section 2.3 (Page 7), and relevant experimental results analysis on this issue will be included after Section 4.3 (Page 18).
2. Regarding the comparison between CARAFE upsampling and other upsampling methods, the following modifications will be made in this paper: a comparative analysis between CARAFE upsampling and other upsampling techniques will be added to Section 2.5 (Page 9), and an experimental result analysis on the upsampling issue will be included in Section 4.3 (Page 18).
3. Addressing the concern of potential errors using morphological algorithms. In Section 2.4 of this paper, we elaborate on the operational mechanism of the morphological algorithm, encompassing its capability to eliminate holes, convexities, concavities, and speckles.  Additionally, we will make the following revision: introducing a supplementary section, 4.2.2 (Page 14), immediately following Section 4.2, to provide an in-depth analysis of the experimental process and outcomes achieved by employing the morphological algorithm on the initial extraction results.
4. Verification of the Model's Generalization Capability.  In this paper, we have curated a comprehensive railway segmentation dataset by collecting aerial images of train stations from various cities and regions across China. This diverse dataset includes samples captured under different weather conditions, terrains, and environmental contexts, effectively encompassing railway station scenarios from a wide geographical coverage of China. Furthermore, to evaluate the model's ability to generalize to other traffic conditions, we performed an in-depth analysis using the publicly available DeepGlobe dataset in Chapter 4. To underscore the robustness of our model's generalization capabilities, we will refine the discussion in Chapter 5 (Page 18) to include a detailed account of the conducted experiments and methodologies employed to validate the model's ability to adapt across varying scenarios and traffic environments.
5. We have resized and adjusted the resolution of Figures 8 and 9, and thoroughly inspected the other figures in this paper to ensure enhanced visual clarity and readability.

 

We sincerely appreciate the opportunity provided by your esteemed journal to enhance the quality of our work. The feedback from respected reviewers like yourself is immensely valuable to our research.

Attached herewith is the revised version of our paper, incorporating all the necessary modifications based on your valuable comments. We hope these revisions meet your expectations and effectively address the concerns raised during the review process. (To facilitate your review of the changes, we have used blue text to highlight the modifications. The red text is revised based on suggestion from another reviewer.)

Once again, we express our gratitude for your feedback. We eagerly look forward to receiving further guidance and suggestions from you to further improve our work.

 

Best regards,

Zuochuang Li

Author Response File: Author Response.docx

Reviewer 2 Report

A Railway Track Extraction Method Based on Improved 3 DeepLabV3+ 

Yanbin Weng, Zuochuang Li, Xiahu Chen, Jing He, Fengnian Liu, Xiaobin Huang, and Hua Yang

 

I have carefully reviewed your paper titled "A Railway Track Extraction Method Based on Improved DeepLabV3+" submitted to the journal "electronics," and I appreciate the effort put into this research. However, I have some observations and questions that I believe would benefit the readers and improve the overall quality of your paper.

1. I find it surprising that this paper was submitted to the "electronics" journal. While your work involves using aerial images and a DeepLabV3+ model, the application in railway track extraction might not directly align with the journal's focus. Please provide a more compelling justification for the relevance of your work to electronics.

2. I would like to see a clearer explanation of how the proposed track extraction method contributes to mapping the railway network. Since OpenRailwayMap already provides shapefiles, readers need to understand the advantages of using your approach. Highlight specific scenarios or use cases where your method outperforms existing solutions and how it aids in railway infrastructure management.

3. The justifications regarding misclassifications are not entirely convincing. Railway tracks indeed have a different landscape compared to highways, but the presence of bumps and holes is usually considered dangerous to operations and is not tolerated. To enhance the credibility of your findings, consider performing an in-depth analysis to identify the primary reasons for misclassifications, such as data limitations or labeling errors.

4. The methodology section mentions dividing the data into training and validation sets, whereas the experimental part refers to dividing it into training and testing datasets. Please clarify this discrepancy to avoid confusion for the readers.

5. Elaborate on the purpose of data augmentation and how it improves the model's performance. To better understand its effectiveness, including a comparison of results achieved with and without augmentation to demonstrate its contribution to the outcomes.

6. Please provide English translations for any Chinese text present in your figures. This will ensure that readers who are not familiar with the language can comprehend the content without any difficulty.

7. The authors must consider sharing your code and instructions with the reviewer to enable them to test the reproducibility of your results. This will strengthen the credibility of your findings and facilitate further research in the community.

8. Explain the significance of using encoders in your model architecture. Describe how they contribute to the track extraction process and why their inclusion is crucial for achieving accurate results.

9. Figures 7 and another figure seem to display roads from the DeepGlobe dataset, which is related to road extraction, not railways. Please clarify the purpose of including these figures and their relevance to your study.

10. Provide a detailed description of the loss function used in your model for the binary classification task. Additionally, share insights into any experiments conducted with alternative activation functions and loss functions to analyze their sensitivity and impact on model performance.

11. Justify the choice of a complex architecture for a seemingly simple classification task. Explain why a simpler model architecture might not be sufficient and highlight the specific aspects of your approach that necessitate the complexity.

I believe addressing these points will significantly improve the clarity, relevance, and impact of your paper.

Author Response

Dear Reviewer,

Firstly, I would like to express my heartfelt gratitude for your review of our submitted paper and providing valuable revision suggestions. We greatly appreciate your diligent examination, as it is crucial for further refining and enhancing the quality of our manuscript. Following the guidance of your suggestions, we have meticulously revised the paper, and below, we provide point-by-point responses to your revision recommendations:

 

1. The rationale behind our choice of the Electronics journal is as follows: Firstly, Electronics journal covers a wide range of topics in electronic engineering and applied electronics, with an openness to accept articles in fields like deep learning and image processing. Secondly, while our research primarily focuses on railway track extraction, the core methodology heavily relies on deep learning and image processing techniques. Beyond the dataset utilized, the majority of our work revolves around exploring and advancing deep learning technologies, aligning perfectly with our team's primary research direction in this domain. Moreover, during our literature review, we have actively sought and studied relevant articles in the Electronics journal that directly relate to our research area.Lastly, the Electronics journal is held in high esteem internationally, known for its influence, reputation, and efficient and unbiased peer-review process. Considering these significant factors, our decision to submit to the Electronics journal was carefully deliberated and based on a thorough examination of the journal's scope and contributions to the field.
2. Regarding the role of our approach in facilitating the creation of railway electronic maps, we provide the following explanation: Firstly, our project's primary objective is the automated generation of station maps, ensuring enhanced safety and efficiency in train scheduling. Currently, companies primarily rely on manual map creation using ArcGIS (Arc Geographic Information System) for this purpose. Our aim is to explore automated methods for map generation, thereby reducing the manual workload significantly.Secondly, unlike OpenRailwayMap, which relies on shapefiles, our approach capitalizes on automation and deep learning techniques.  This allows us to efficiently process large-scale railway data, reducing the manual effort needed for map updates and maintenance. Furthermore, our method enables swift and frequent updates, which are of utmost importance in dynamic station railway environments. Lastly, it is crucial to note that the extraction of railway tracks merely represents the initial step in generating station maps. The subsequent stages involve processing other critical railway elements such as turnouts, signals, and embankments until we obtain station maps encompassing various railway features. These specific locations of railway features cannot be accurately marked using OpenRailwayMap.
3. In order to address potential sources of errors in deep learning method, we will incorporate the following modification: Prior to introducing the innovations of this paper on page 2 of Chapter 1 (Page 2), we will insert a paragraph that provides a comprehensive analysis of the factors contributing to possible inaccuracies.
4. We sincerely apologize for the discrepancy in the descriptions of the datasets.  During our work, we utilized three distinct datasets, namely the 'Training Set,' 'Test Set,' and 'Validation Set,' with the 'Test Set' and 'Validation Set' using the same subset. We deeply regret any confusion caused by our unclear language. To rectify this, we have revised the description of the dataset partitioning in Section 3.1(Page 10) of Chapter 3 to ensure the most straightforward comprehension for our readers.
5. Regarding the role and effectiveness of data augmentation, this paper proposes the following modification: In Chapter 4, a new section, 4.2.1(Page 14), will be added to provide a detailed description of the impact of data augmentation on the initial extraction results.
6. Thank you very much for your reminder. The presence of Chinese characters in the figures and charts was our oversight. Currently, all the text in the figures and charts of the paper has been translated into English.
7. Thank you very much for your suggestion. We understand the importance of sharing code and documentation. However, the project is still under active development, and this code remains a part of ongoing research efforts. Sharing it at this stage could potentially hinder future development and competitive advantage. Once the project is concluded and no longer subject to confidentiality, we will consider uploading the code and documentation to Github for further research purposes.
8. In regard to the role of the encoder, the following modification will be made in this paper: A new paragraph will be added in Chapter 2, Section 2.2 (Page 4), to provide an analysis of the functions of both the encoder and the decoder.
9. Our rationale for introducing the DeepGlobe dataset in the paper is as follows: Our research primarily focuses on railways, but obtaining a large amount of foreign railway aerial imagery data, especially from unmanned aerial vehicles (UAVs), may not be readily available due to limited accessibility. As a result, conducting extensive generalization experiments with foreign railway datasets becomes challenging. Although the DeepGlobe dataset is originally intended for road extraction, its underlying principles for extraction are similar to railways, making it a suitable alternative for evaluation purposes. Additionally, the dataset is substantial in size and easily accessible, thus making it an appropriate choice for assessing the model's generalization capabilities.
10. For a detailed explanation of the loss function, the following modification will be made in this paper: A comprehensive introduction of the loss function will be included in Chapter 3, Section 3.2 (Page 11), along with an analysis of its impact on the model's performance.
11. The reasons for selecting complex architectures for what might seem like a simple task are as follows: Firstly, it is attributed to the nature of the data. The dataset consists of railway track images captured from UAV aerial views, and inherent noise is unavoidable. Complex architectures excel in learning underlying patterns behind the noise, enabling better handling of such situations and ultimately improving accuracy. Secondly, the task demands precise boundary information extraction for railway tracks. Simple architectures may struggle to capture these complexities, leading to blurry or inaccurate segmentations at object boundaries. In contrast, complex architectures leverage multi-level feature representations, facilitating better capturing of object shapes and boundary information. Lastly, achieving accurate semantic segmentation necessitates considering contextual information around pixels to better comprehend their semantic classes. Complex architectures with larger receptive fields and enhanced feature extraction capabilities are capable of capturing broader contextual information, thus enabling more robust semantic inference for each pixel.

 

Thank you once again for your careful review of our paper and valuable suggestions. We have put forth our utmost effort in making the necessary revisions, and we hope our responses meet your requirements. Attached herewith is the revised version of our paper, incorporating all the necessary modifications based on your valuable comments. If there are any other aspects that need improvement, please kindly advise, and we will promptly make the necessary adjustments. (To facilitate your review of the changes, we have annotated the modified areas in the attached document with red text. The blue text is revised based on suggestion from another reviewer.)

Lastly, we sincerely appreciate your support and assistance. Under your guidance, we believe our paper will be further enhanced. We look forward to receiving your final approval.  Thank you!

 

Best regards,

Zuochuang Li

Author Response File: Author Response.docx