Comparative Study of Lightweight Deep Semantic Segmentation Models for Concrete Damage Detection
Round 1
Reviewer 1 Report
The article is devoted to the actual problem of the modern society - the effect of Deep semantic segmentation models for concrete damage detection. The modern estimation methods were applied in the analysis. The article has a deep scientific and applicable industrial soundness and can be accepted for publication. As a wish for article improvement - it could be better if the authors presented the well-know methods of the problem determination in their study area and compare with the method used in the article.
The Figures 8-10 is heavy to understand, because there is no description on the wavelengths or any indicators describing the axis labels.
Author Response
First of all, the authors appreciate the efforts of the reviewers. In this response, the authors did their best to respond all questions given by the reviewers. The changed parts made according to the response is marked red in the revised manuscript.
C1. The article is devoted to the actual problem of the modern society - the effect of Deep semantic segmentation models for concrete damage detection. The modern estimation methods were applied in the analysis. The article has a deep scientific and applicable industrial soundness and can be accepted for publication. As a wish for article improvement - it could be better if the authors presented the well-known methods of the problem determination in their study area and compare with the method used in the article.
A1. This study compare 4 different semantic segmentation models by benchmarking the performance of the DeepLabV3+ as shown in Table 1. DeepLabV3+ is a well-known state of the art deep learning model widely used for semantic segmentation task. Therefore, this article already compares the performance of other models with DeepLabV3+. The conventional well-known method would be a visual inspection of damages from the images, and this article calculates the performance based on the ground-truth of damage area obtained by visual inspection of the test images.
C2. The Figures 8-10 is heavy to understand, because there is no description on the wavelengths or any indicators describing the axis labels.
A2. Figures 8-10 are improved and merged the ground truths and predictions of all models with original raw image for better understanding.
Reviewer 2 Report
The original contribution of the paper must be more clearly mentioned.
Author Response
First of all, the authors appreciate the efforts of the reviewers. In this response, the authors did their best to respond all questions given by the reviewers. The changed parts made according to the response is marked red in the revised manuscript.
C1. The original contribution of the paper must be more clearly mentioned.
A1. The contribution of the paper is presented in the bullet format from line 140 to line 153 of Introduction in the revised manuscript as below:
In summary, the main contributions of the proposed study are as follows:
- Three lightweight models (ENet) [42], context-guided network (CGNet) [45], and efficient symmetric network (ESNet) [46]) and two heavyweight models (deep dual-resolution networks (DDRNet) [47] and DeepLabV3+ [48]) were compared for the damage segmentation from the structural images to investigate the best model that can be embedded in the edge computing devices with less computational power.
- A concrete dataset that contains four types of concrete damage, i.e., cracks, efflorescence, spalling, and rebar exposure was constructed for training and testing of semantic segmentation models. Images were collected from online and real concrete structures in South Korea.
- The lightweight segmentation models were benchmarked for the detection of multiple types of concrete damage, and the tradeoff between the number of model parameters and accuracy has been investigated.
Reviewer 3 Report
There are no comments. The article is interesting and informative.
Author Response
The authors really appreciate the positive review result about the submission.
Reviewer 4 Report
This manuscript compared and analyzed the performance of five semantic segmentation models that can be used for damage detection. The manuscript is novel and well written. There are some suggestions presented below to improve the scientific impact of the manuscript:
- The abstract is too long, it can be summarized.
- The introduction can be reduced.
- The quality of all figures should be increased.
- Are the Figures 5, 6 and 7 depicted by the authors?
- No information regarding the suggested models were provided. Please clarify how the other authors can reuse and reformulate the proposed models? In machine learning techniques the process of using the proposed models is important. For instance, the linking weights and biases in ANN, the closed form equation in GP and GEP and GMDH models should be provided. There are some suggestions to be read and be referenced as below:
· Seydi, S. T., Rastiveis, H., Kalantar, B., Halin, A. A., & Ueda, N. (2022). BDD-Net: An End-to-End Multiscale Residual CNN for Earthquake-Induced Building Damage Detection. Remote Sensing, 14(9), 2214.
· Daneshvar, M. H., Saffarian, M., Jahangir, H., & Sarmadi, H. (2022). Damage identification of structural systems by modal strain energy and an optimization-based iterative regularization method. Engineering with Computers, 1-21.
· Hajializadeh, D. (2022). Deep-learning-based drive-by damage detection system for railway bridges. Infrastructures, 7(6), 84.
- The conclusion can be summarized and be presented in bullet format.
Author Response
First of all, the authors appreciate the efforts of the reviewers. In this response, the authors did their best to respond all questions given by the reviewers. The changed parts made according to the response is marked red in the revised manuscript.
C1. The abstract is too long, it can be summarized
A1. Abstract is summarized in the revised manuscript and changes can be found in the revised manuscript from line 16 to line 20.
C2. The introduction can be reduced.
A2. Some of the less important sentences from the introduction section is remove. These sentences in the revised manuscript can be found in lines 41, 42, 43, 44, 50, 51, 63, 76, 77, 81, 82, 99,100, 117 and 118 respectively.
C3. The quality of all figures should be increased.
A3. The quality of all figures is improved in the revised manuscript.
C4. Are the Figures 5, 6 and 7 depicted by the authors?
A4. Figures 5, 6 and 7 shows the architecture of the models adopted for this study. These figures are reconstructed and modified by the authors for better understanding.
C5. No information regarding the suggested models were provided. Please clarify how the other authors can reuse and reformulate the proposed models? In machine learning techniques the process of using the proposed models is important. For instance, the linking weights and biases in ANN, the closed form equation in GP and GEP and GMDH models should be provided.
A5. We changed manuscript in Section 3.2 (Training details) to convey the readers how the models were trained in detail. The segmentation models were proposed based on the convolutional neural networks, whose filters (or kernels) are to be trained, which is somewhat different from the training of common machine learning models such as ANN and GP. The filters of convolutional layers in the first four models (e.g., ENet, CGNet, ESNet, and DDRNet) were trained from scratch, while the filters in the backbone network (e.g., ResNet-50) of DeepLabv3+ were updated from the filters of the pretrained model using ImageNet dataset. The details of convolutional layers, whose filters are to be trained, used in the models were explained in Section 2. The changes can be found in the revised manuscript in Section 3.2 (Training details) from line 365 to line 374 as bellow:
To train the deep learning model from scratch all the learnable model parameter like weights (convolution filters) initialize randomly according to the input data during training and loss is calculated after each iteration. These model parameters are monitored by loss function and the training stop when the loss get minimize and close to zero. Models like DeepLabV3+ used backbone network (ResNet-50) for classification task and decoder head for semantic segmentation. This backbone network used pretrained weights of ImageNet excluding last fully connected layer and help the model to optimize quickly during training also known as transfer learning. In this study DeepLabV3+ also used pretrained weights in the encoder module ResNet-50 as transfer learning.
C6. The conclusion can be summarized and be presented in bullet format.
A6. The conclusions are summarized and presented in bullet format in the revised manuscript.
Round 2
Reviewer 4 Report
The revised version of the manuscript can be accepted.
