Visual Detection of Road Cracks for Autonomous Vehicles Based on Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a study on road crack detection for autonomous vehicles using deep learning. While the research offers valuable results, there are several weaknesses that need to be addressed to enhance the scholarly contribution of the paper.

(1) The title should directly reflect the focus on visual detection of road cracks for autonomous vehicles based on deep learning.

(2) The comparison of previous road crack detection methods is insightful, but the proposal should effectively compare both domestic and foreign fruit detection methods to address an important limitation. The paper's limitations should be articulated more rigorously and realistically.

(3) The document includes a total of 30 references, with 90% published within the last 5 years, 2% published within the last 5-10 years, and 7.5% without a publication year indicated. The number of references is inadequate.

(4) The article's approach of modifying and overlaying existing neural networks and increasing training frequency lacks innovation.

(5) The abstract is comprehensive and well-structured, but should include numerical indicators related to the study's results.

(6) The introduction should provide a broader perspective on the research topic, align the work with previous studies in the field, and clarify the study's novelty and contribution.

(7) The paper adequately presents the proposed detection technology and its objectives. However, the limitations of the work should be rigorously hypothesized and evidenced.

(8) The authors may add more state-of-art application articles for the integrity of the manuscript (3D vision technologies for a self-developed structural external crack damage recognition robot; Automation in Construction. Enhanced Precision in Dam Crack Width Measurement: Leveraging Advanced Lightweight Network Identification for Pixel-Level Accuracy; International Journal of Intelligent Systems.).

(9) Clarify the specific meaning of the Kappa coefficient in line 447 and its impact on prediction accuracy.

(10) The technology's high requirements for on-site fruit shooting work and potential errors during shooting should be further addressed.

(11) Further research scope and the practical implications of the research should be discussed.

(12) The document should include the limitations related to damages and strengthen this section for a more thorough assessment.

Author Response

Authors’ Responses to Reviewers’ Comments

 

We thank all the reviewers for their constructive suggestions and insightful comments. According to their comments, we improve the quality of this manuscript significantly and try to address the reviewers’ concerns. In the following, we list each reviewer’s comments, and immediately after which, we provide our corresponding responses, along with the description of the corresponding change(s) that we have implemented in the revised manuscript.

 

Reviewer #1

 

Overall Comment:

 

This paper presents a study on road crack detection for autonomous vehicles using deep learning. While the research offers valuable results, there are several weaknesses that need to be addressed to enhance the scholarly contribution of the paper.

Reply:

Many thanks. We greatly appreciate the positive comments. In the following, we try to address your concerns and improve the quality of this paper. Thank you for your help.

 

• The title should directly reflect the focus on visual detection of road cracks for autonomous vehicles based on deep learning.

 

Author Response: Thank you for pointing this out. We have revised the title. Please see Page 1 line 1.

 

 

• The comparison of previous road crack detection methods is insightful, but the proposal should effectively compare both domestic and foreign fruit detection methods to address an important limitation. The paper's limitations should be articulated more rigorously and realistically.

 

Author Response: Thank you for pointing this out. Our proposed work primarily focuses on road crack detection, which stands as the central goal of our research endeavor. While there are existing methods for detecting various types of fruits, both domestically and internationally, we must emphasize that fruit detection is not within the scope of our investigation. Unfortunately, we are unable to demonstrate our findings in this area, as our research is specifically tailored towards addressing the challenges associated with road crack detection. While we acknowledge the importance of fruit detection methodologies, our work is distinctively centered on enhancing the accuracy and efficiency of crack detection algorithms for concrete pavement structures.

 

• The document includes a total of 30 references, with 90% published within the last 5 years, 2% published within the last 5-10 years, and 7.5% without a publication year indicated. The number of references is inadequate.

 

Author Response: Thank you for pointing this out. We have revised the references; in which publication year missing. We have also added new references to improve our introductions.

 

• The article's approach of modifying and overlaying existing neural networks and increasing training frequency lacks innovation.

 

Author Response: Thank you for pointing this out. The paper proposed methodology delves into the crucial task of road crack detection, emphasizing its vital role in assessing the structural integrity of concrete pavements, particularly in contexts such as autonomous vehicle navigation systems. It points out the inherent challenges associated with traditional image-based methods, which often require extensive preprocessing to extract relevant crack features, posing practical limitations in real-world scenarios where concrete surfaces exhibit diverse forms of noise, such as variations in lighting conditions, surface textures, and environmental debris. In response to these challenges, the research introduces an innovative approach that merges a random forest machine learning classifier with a DCNN architecture. By integrating these two methodologies, the proposed hybrid model leverages the strengths of both techniques, allowing for more robust and accurate crack detection. Notably, the paper meticulously explores various hyperparameters and experimental configurations (Ablation Study) to identify an optimal base learning rate of 0.001, crucial for effective model training. This fine-tuned learning rate significantly contributes to the model's performance, culminating in an impressive maximum validation accuracy of 99.97%. This high level of accuracy underscores the efficacy and reliability of the hybrid approach in detecting road cracks with precision and consistency, even in challenging real-world environments.

 

• The abstract is comprehensive and well-structured, but should include numerical indicators related to the study's results.

 

Author Response: Thank you for pointing this out. We have added numerical indicators such as test accuracy, precision, recall, and f1-score to present proposed model results in Abstract. Please reads as follows on Page 1 line 18-21.

 

 

• The introduction should provide a broader perspective on the research topic, align the work with previous studies in the field, and clarify the study's novelty and contribution.

 

Author Response: Thank you for pointing this out. We have revised our introduction and add new reference to broader perspective on the research topic.

 

• The paper adequately presents the proposed detection technology and its objectives. However, the limitations of the work should be rigorously hypothesized and evidenced

 

Author Response: Thank you for pointing this out. We have added the limitation of proposed work, as these limitations are focus points of our future research. Please read as follow; Page 20 line 762-770.

 

• The authors may add more state-of-art application articles for the integrity of the manuscript (3D vision technologies for a self-developed structural external crack damage recognition robot; Automation in Construction. Enhanced Precision in Dam Crack Width Measurement: Leveraging Advanced Lightweight Network Identification for Pixel-Level Accuracy; International Journal of Intelligent Systems.).

 

Author Response: Thank you for pointing this out. We have improved our related work with more state-of-art application articles. Please reads as follows on Page 5 line 207-212, and Page 4 line 189-196.

 

• Clarify the specific meaning of the Kappa coefficient in line 447 and its impact on prediction accuracy.

 

Author Response: Thank you for pointing this out. We have added the details description about specific meaning of the Kappa coefficient, and what its impact on prediction on our proposed model prediction. Please reads as follows on Page 13 line 545-550.

 

 

• The technology's high requirements for on-site fruit shooting work and potential errors during shooting should be further addressed.

 

Author Response: Thank you for pointing this out. Our proposed work primarily focuses on road crack detection, which stands as the central goal of our research endeavor. While there are existing methods for detecting various types of fruits, both domestically and internationally, we must emphasize that fruit detection is not within the scope of our investigation. Unfortunately, we are unable to demonstrate our findings in this area, as our research is specifically tailored towards addressing the challenges associated with road crack detection. While we acknowledge the importance of fruit detection methodologies, our work is distinctively centered on enhancing the accuracy and efficiency of crack detection algorithms for concrete pavement structures.

 

• Further research scope and the practical implications of the research should be discussed.

 

Author Response: Thank you for pointing this out. We have added the details description about practical implications of our proposed model. Please reads as follows; Page 20 line 744-748. Furthermore, we have highlighted main scope of this study in Page 3 line 120-132.

 

• The document should include the limitations related to damages and strengthen this section for a more thorough assessment.

 

Author Response: Thank you for pointing this out. We have added the details description about limitations and strengthen of our proposed model. Please read as follows: Page 20 line 750-770.

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The present study introduces a novel hybrid methodology for road information detection within autonomous vehicles, integrating both positive and negative image datasets. The proposed approach amalgamates a Random Forest classifier with transfer learning methodologies. Specifically, this investigation exploits layers derived from three pre-trained models, namely Xception, InceptionV3, and MobileNet. This fusion of pre-trained models with Random Forest serves to mitigate overfitting while simultaneously augmenting the efficacy of high-dimensional feature extraction, thereby facilitating enhanced performance.

·         The paper’s organization is perfect.

·         The quality of the writing is excellent.

·         The presentation methodology (i.e., methods, experiments, and analysis) is clear and understandable.

·         The suggested approach's performances are well assessed using several experiments and comparisons.

·         The obtained results are acceptable, competitive, and surpass all compared works.

·         The references are recent and of good quality.

However, before accepting the manuscript, I would like that authors consider the following points:

·         The paper’s organization should be added in the last part of the introduction.

·         The authors should use at least two datasets to guarantee the generalization of the proposed method.

·         References should be written with MDPI format.

Author Response

Authors’ Responses to Reviewers’ Comments

 

We thank all the reviewers for their constructive suggestions and insightful comments. According to their comments, we improve the quality of this manuscript significantly and try to address the reviewers’ concerns. In the following, we list each reviewer’s comments, and immediately after which, we provide our corresponding responses, along with the description of the corresponding change(s) that we have implemented in the revised manuscript.

 

Reviewer #2

 

Overall Comment:

The present study introduces a novel hybrid methodology for road information detection within autonomous vehicles, integrating both positive and negative image datasets. The proposed approach amalgamates a Random Forest classifier with transfer learning methodologies. Specifically, this investigation exploits layers derived from three pre-trained models, namely Xception, InceptionV3, and MobileNet. This fusion of pre-trained models with Random Forest serves to mitigate overfitting while simultaneously augmenting the efficacy of high-dimensional feature extraction, thereby facilitating enhanced performance.

• The paper’s organization is perfect.
• The quality of the writing is excellent.
• The presentation methodology (i.e., methods, experiments, and analysis) is clear and understandable.
• The suggested approach's performances are well assessed using several experiments and comparisons.
• The obtained results are acceptable, competitive, and surpass all compared works.
• The references are recent and of good quality.

However, before accepting the manuscript, I would like that authors consider the following points:

Reply:

Many thanks. We greatly appreciate the positive comments. In the following, we try to address your concerns and improve the quality of this paper. Thank you for your help.

• The paper’s organization should be added in the last part of the introduction.

Author Response: Thank you for pointing this out. We have added the details description about paper’s organization. Please read as follows; Page 3 line 135-140.

• The authors should use at least two datasets to guarantee the generalization of the proposed method.

Author Response: Thank you for pointing this out. We have used multiple datasets in our upcoming paper as this is the future part of this paper. Please Read Page 21 Line 793-797. Additionally, we are utilizing a diverse range of evaluation metrics to ensure the robustness and applicability of the proposed method in this work.

• References should be written with MDPI format.

Author Response: Thank you for pointing this out.  We have revised References to MDPI format.  Please read as follows: Page 22-24.

 

 

 

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

After a review of the current methods for crack detecting in concrete After a review of current methods for crack detection in concrete pavements, the combination of a random forest machine learning classifier with a deep convolutional neural network is tested with a large database.

The article is well written. The test results are well represented. The weakness consists above all in the lack of information on the processing time required by the various methods, which is an important element.

Author Response

Authors’ Responses to Reviewers’ Comments

 

We thank all the reviewers for their constructive suggestions and insightful comments. According to their comments, we improve the quality of this manuscript significantly and try to address the reviewers’ concerns. In the following, we list each reviewer’s comments, and immediately after which, we provide our corresponding responses, along with the description of the corresponding change(s) that we have implemented in the revised manuscript.

 

Reviewer #3

 

Overall Comment:

 

Comments and Suggestions for Authors

After a review of the current methods for crack detecting in concrete After a review of current methods for crack detection in concrete pavements, the combination of a random forest machine learning classifier with a deep convolutional neural network is tested with a large database.

The article is well written. The test results are well represented.

Reply:

Many thanks. We greatly appreciate the positive comments. In the following, we try to address your concerns and improve the quality of this paper. Thank you for your help.

 

• The weakness consists above all in the lack of information on the processing time required by the various methods, which is an important element.

Author Response: Thank you for pointing this out. We have added the details about time required to base and our proposed model. Please read as follows: Page 12 line 519-523.

 

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

accept