Fast Image Classification for Grain Size Determination

Round 1

Reviewer 1 Report

A very interesting article is presented. The relevance of the work is beyond doubt. The proposed methods and approaches are of scientific and technical importance. The results of the work allow us to give an assessment.
However, it would be interesting to investigate the developed neural network using noisy images, images with flares, images with blur, low contrast, etc.

Author Response

A very interesting article is presented. The relevance of the work is beyond doubt. The proposed methods and approaches are of scientific and technical importance. The results of the work allow us to give an assessment.
However, it would be interesting to investigate the developed neural network using noisy images, images with flares, images with blur, low contrast, etc.

Response: Thank you for the constructive suggestions. Since the grain size images we obtained are all taken under the microscope, there are no blur and out-of-focus conditions at a fixed focal length. However, some images we use are more or less having around 10% noises (such as oxides and inclusions), but will not affect the final result.

Reviewer 2 Report

In this paper, the authors propose the CNN-based FIC model and use it to determine grain size for carbon steel. the authors use a real metallographic dataset to compare FIC with other deep learning network architectures. The experimental results show that the proposed method yields a classification accuracy of 99.7%.

I recommend the acceptance of this paper with minor revisions. A suggestions are raised to further improve the manuscript.

It is better to collect a larger dataset to further evaluate the effectiveness of FIC in determining grain size.

Author Response

In this paper, the authors propose the CNN-based FIC model and use it to determine grain size for carbon steel. the authors use a real metallographic dataset to compare FIC with other deep learning network architectures. The experimental results show that the proposed method yields a classification accuracy of 99.7%.
I recommend the acceptance of this paper with minor revisions. A suggestions are raised to further improve the manuscript.
It is better to collect a larger dataset to further evaluate the effectiveness of FIC in determining grain size.

Response: Thank you for the constructive suggestions. However, because the microstructure image of steel is not easy to obtain, the most effective method is to use the data argumentation method mentioned in our manuscript. In the future, we will use the images obtained from each experiment to increase our database.

Reviewer 3 Report

The manuscript addresses a very interesting and up-to-date topic in steel research. Grain size is one key factor influencing material quality. Data quality and amount are essential aspects for the successful application of deep learning approaches. Especially considering this matter, some aspects need to be explained more clearly:

1. Grain size data: Information about the used grain sizes is missing. Grain size can vary significantly from sample to sample (depending on steel composition and production state). Which range of grain sizes has been analysed? What was the mean value and standard deviation of the grain sizes?
2. From which production step the samples were taken? Are these industrial samples? Which metallographic procedure (etching method) has been used for sample preparation?
3. Amount of data: If I understood correctly, data has been augmented artificially by different modifications of the data set. However, this not change the mean grain size distribution or? So isn't this critical in terms of the applicability of the proposed FIC approach? Which accuracy can be obtained if the grain size distribution is wider?
4. More information regarding the accuracy of the model should be provided. E.g. a confusion matrix pointing out the problem of classification using the VGG16 method in comparison to FIC model. What is the reason for the limitations of other models?
5. Literature: More literature regarding "classical" austenite grain size determination should be included in the paper, e.g. https://doi.org/10.2355/isijinternational.39.271 ,, also considering some recent ones e.g.: https://doi.org/10.1016/S1044-5803(01)00142-5 , https://doi.org/10.1016/j.mtcomm.2021.102468

Author Response

The manuscript addresses a very interesting and up-to-date topic in steel research. Grain size is one key factor influencing material quality. Data quality and amount are essential aspects for the successful application of deep learning approaches. Especially considering this matter, some aspects need to be explained more clearly:

1. Grain size data: Information about the used grain sizes is missing. Grain size can vary significantly from sample to sample (depending on steel composition and production state). Which range of grain sizes has been analyzed? What was the mean value and standard deviation of the grain sizes?

Response: There are two phases of grain size analyzed in this study, one is ferrite, and the other is austenite. As described in the last sentence of the first paragraph on page 8, for the grain size distribution of ferrite phase was in the range of grade 7 and 10 with ASTM E112; likewise, for austenite was in the range of grade 5 to 8. As so far, the mean value for grain size in a metallographic picture could be analyzed and applied in quality control in industry. The deviation of grain size will be studies in the future.

2. From which production step the samples were taken? Are these industrial samples? Which metallographic procedure (etching method) has been used for sample preparation?

Response: All metallographic samples were taken from the hot-rolled rods for carbon steel of China steel corporation. The microstructure features of the cross sections for the rods would be revealed after cutting, mounting, grinding, polishing and finally using the chemical etching with the natal solution which is mixtures of ethanol and nitric acid.

3. Amount of data: If I understood correctly, data has been augmented artificially by different modifications of the data set. However, this not change the mean grain size distribution or? So isn't this critical in terms of the applicability of the proposed FIC approach? Which accuracy can be obtained if the grain size distribution is wider?

Response: We thank the reviewer for the constructive suggestions. In our manuscript, we used the most common three data augmentation techniques for grain size images: flipping, cropping, and rotation, each of which is associated with two parameters. Those methods will not cause the deformation of the grain size images and the change of the size distribution. Therefore, our proposed method will not be affected by the data augmentation method.

4. More information regarding the accuracy of the model should be provided. E.g. a confusion matrix pointing out the problem of classification using the VGG16 method in comparison to FIC model. What is the reason for the limitations of other models?

Response: Yes, I completely agree with the reviewer’s opinions. The experimental results show that FIC is the fastest classifier and VGG16 is the slowest, mostly because VGG16 has 138 million parameters, leading to greater computational costs. In the DenseNet201 and ResNet50, the main network architectures of feature extraction are obviously not enough for image classification, resulting in lower accuracy. In addition, although Darknet53 has the highest accuracy on the training and testing sets, its BFlops of 56.88 is unacceptable for real-time image classification. The discussion of the issue is added to section 4.3 of the manuscript.

5. Literature: More literature regarding "classical" austenite grain size determination should be included in the paper, e.g. https://doi.org/10.2355/isijinternational.39.271 ,, also considering some recent ones e.g.: https://doi.org/10.1016/S1044-5803(01)00142-5 , https://doi.org/10.1016/j.mtcomm.2021.102468

Response: Thank you for the constructive suggestions. Yes, I completely agree with the reviewer’s opinions. Those recent studies [3-5] are discussed in Section 2 of the manuscript.

Round 2

Reviewer 3 Report

Thank you for your clarification of the mentioned points. The manuscript is now clear and concise. Can be published in the present form.