Efficient Hair Damage Detection Using SEM Images Based on Convolutional Neural Network

Round 1

Reviewer 1 Report

Article prepared correctly. An interesting tool was used to estimate the quality of hair. What is worth noting is that authors created a hair microscopy dataset from male and female hair data samples and divided the data into weak damage, damage and high damage according to the degree of hair damage. A new convolutional network based on a spatial attention module was proposed and verified with the previously created dataset. It was also proved that the network model is efficient and robust to disturbances.

Author Response

We have revised the grammatical errors in the paper and supplemented and improved the experimental part.

Reviewer 2 Report

In this manuscript, titled “Efficient hair damage detection using SEM images based on convolutional neural network”, Man et al. created a data base containing SEM images of hair, to detect and assess hair damage with the help of a convolutional neural network. To that end, a channel and spatial attention mechanism was implemented in the Residual Channel Spatial Attention Network.

 

General remarks

• There is no information about the collection and classification of hair samples in the methods section, and very little information in the results section. How many samples have been collected and analyzed in total, and what properties did they have (hair color, use of hair dye, etc.)? What are “mixed men and women”? Why are there conflicting statements about the age groups included in the study (20-60 vs 18-55 age group)?
• The authors state that “the collection of hair SEM microscopic images is complicated and expensive, it is impossible to collect data on a large scale.” I do not necessarily agree with this statement, since dozens of hairs can easily be mounted on a single SEM stub and can be imaged in minutes with the help of a modern SEM. The authors should at least disclose how many individual hair samples have been used to create the database; the way the “data enhancement algorithm” is described, the authors could have theoretically used a single SEM image per class.
• The authors show SEM images of hair with alleged “weak damage”, “damage”, and “high damage”. What are the criteria for the assessment and classification of hair damage? There are no negative or positive controls included, for example hair know to be damaged by chemical stress, mechanical stress, etc. Under these circumstances, how can we know that this classification is correct and representative of hair stress, rather than other differences in hair morphology?
• The purpose of the discussion is to discuss and interpret the results in the light of previous studies, yet there is not even a single reference mentioned, completely missing the point of the discussion section!

 

Specific comments

• Abbreviations are not explained in the manuscript, for instance ReLU (Figure 2), C, H, W (Figure 5), or the factors in formulas 1 and 2.
• The writing of this manuscript is extremely sloppy and contains not only a plethora of grammatical errors, but also many nonsensical sentences. I do not have the time nor is it my job to correct the manuscript for the authors, thus I will only highlight a few nonsensical passages as examples in the following.

Lines 35-39: Exposure to various physical and chemical stresses (…) is of great significance for the diagnosis of skin diseases (…)

Lines 87-88: It is worth noting that Kim et al. A hair damage classification  system  was  developed.

Lines 90-92: The grading system depends on the scan Electron microscope (SEM) images are based on (…)

Line 97: In these studies, most of the studies (…)

Lines 157-161: We proposed RCSAN network as show as in Figure3 is based on the residual network [16].  the  basis  of,  We  deal  with  complex  and  difficult-to-detect  hair  damage  feature  regions by stacking multiple residuals and a residual attention block formed by combining attention blocks, thereby establishing an efficient network framework for hair damage detection. as show as Fig.3.

Author Response

1. Modified writing and grammatical errors.
2. The hair SEM data set we created uses 50 males and 50 females. Each tester collects 30 hair samples, and the middle section of the hair is used as the observation area. Our task is to design a model that automatically detects the degree of hair damage, focusing only on the degree of hair damage, so it does not pay attention to the cause of hair damage. We rented SEM equipment for a limited time and were expensive for each use, so we were unable to create a large-scale data set. When we use the SEM equipment, each base uses 30 hair samples for observation. In order to ensure the reliability of the data, the imaging has been fine-tuned many times, and the observation time is 3-4 hours each time.
3. In the second chapter of the article, we introduced the research of others. However, there are almost no models for automatically detecting hair damage based on artificial intelligence.
4. Our hair damage assessment is based on the degree of damage to the stratum corneum on the surface of the hair.

Reviewer 3 Report

The authors submitted a very interesting work, my suggestions to improve it are: 

• Focus of your paper is to propose a new dataset and to report baseline performance on it, I suggest to upload the dataset on a repository as zenodo; 
• Add details on the dataset, e.g. how many men and females have you collected? 
• You randomly split the dataset into training/test sets, you have mixed images of the same person between training and test sets, in this way you obtained biased performance. I suggest a k-fold cross validation where all the images of a given person belong to the training or test set. 
• As shown in table 2 SACN-Net obtains very good performance, add some comments on the obtained result, e.g. Is it enough good for real applications? 
• It is already shown in the literature that combining CNNs based on different topologies permits to boost performance, you could add some fusions (e.g. by sum rule) in table 2. 
• Section 6 is very brief, add more comments and future works.  

Author Response

We have revised the article details and grammar.

We collected hair samples of 50 men and 50 women in the 20-60 age group and used Hitachi S-4700 SEM (Scanning Electron Microscope) to generate hair microscopic images. We created the hair SEM data set that has a total of 15,000 hairs microscopic images of various types, contains:
5000 images of weakly damage hair microscopic.
5000 images of damage hair microscopic.
5000 images of high damage hair microscopic.
select 80% of the data as the training set and 20% of the data as the test set in a random and non-repetitive manner. In the conclusion of Chapter 6, we have supplemented the content of this chapter.

Round 2

Reviewer 2 Report

Regarding language and style, there has been no effort to try and correct the wording of this manuscript apart from the few examples I pointed out in my previous report. Abbreviations remain undefined. Consider having the manuscript proof-read by an english speaker.

Questions from my previous report have not been addressed, the reader is still left to guess the methodical details like how many hairs have been used, what percentages of hair were assigned to each category etc. etc.

The category “damage” should be labelled “moderate damage”.

Due to the lack of controls, we do not know what the significance of the data set is, and if it would be working for a real application. Intentional damaging of hair followed by damage assessment by SEM has been done as early as the 1970s (e.g. J. Soc. Cosmet. Chem. 27, 155-161, 1976), and samples like these could be used to verify that the database is actually detecting hair damage.

Unfortunately, the discussion section is still all but omitted, not due to the lack of literature, but because the results are not critically discussed and shortcomings are not addressed. For instance, it needs to be asked if a categorization into only 3 classes is useful, when other papers have shown that detection of subtle differences in chemical hair damage requires a fine point scale (Lee et al), and the further away from the root a sample is taken, the larger the natural occurring mechanical damage (wear) becomes.

Author Response

We corrected the grammatical problems in the paper and the explanation of the meaning of the abbreviations.
The purpose of our research is to design an automatic and efficient recognition and detection of hair damage based on a convolutional neural network algorithm. How the hair is damaged is beyond the scope of our discussion. The hair damage data set we have established collects daily hair samples from testers and observes and classifies these samples to create data sets of weak damage, moderate damage, high damage, and 3 types of hair damage. The data set is used for the training of the hair damage detection algorithm model we designed. After the model is trained, it is used in scenes such as daily hair damage detection and medical health diagnosis.

Author Response File: Author Response.pdf

Reviewer 3 Report

Revision weel done

Author Response

We have modified the details of the article.

Author Response File: Author Response.pdf