Research on Wear Detection of End Milling Cutter Edge Based on Image Stitching
Round 1
Reviewer 1 Report
This paper introduced a new CNN model for the image segmentation training and testing of the side edge wear area. The new CNN model, named FCN, has a short training and recognition time with fewer parameters required.
I give points for practicality of this model. The FCN can be used for accurately and quickly segmenting and calculating the wear value of the side edge of the end mill.
Required minor correction:
The author needs to describe a scenario why FCN is practical and needed in real world. I can understand why it is needed, but it was not that clear from the introduction.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
A brief summary
The authors address the important topic of automated monitoring and assessment of cutting edge wear on a cutting machine tool. The study used a vision-based method to collect a series of images of the edge under investigation, which were combined in post-processing to produce a single image of the edge. To extract the edge and determine the maximum wear value of the end mill, the authors used the fully convolutional neural network (FCN) segmentation 14 algorithm of the VGG16 structure. The results of the FCN were compared with the results of calculations for the K-means and Qtsu threshold methods.
General comments
The main problem with the article is that it is "sparing" in research details. In this connection, I have three main points:
1) Collecting the cutter end mill photos:
1.1) how did this process go - first 10 minutes of milling with an unused cutter, then collecting photos with different spindle speeds?
1.2) were the frames per minute the same at different spindle speeds?
2) FCN:
2.1) what is the VGG16 structure and why was this one chosen (have other FCN structures been studied)?
2.2) the biggest unknown for me is: what the training data looked like and what were the input parameters of the FCN learning process (details)?
2.3) how were 3,300 edge images obtained from 30 end mills? The authors do not comment in any way on such a number of inputs. It would be worth referring to the literature with an explanation if there is a sufficient number of teaching images for this type of network.
2.4) In the context of detecting edge damage and the degree of tool wear, it would be worth considering to what extent such high compression of the input image (from 9792x2048 to 1070x224) resulted in the loss of key wear information, and whether image compression had an impact on the results of other methods (K- average and Qtsu threshold). It is not clear whether the results of VB calculations for K-means and Qtsu threshold in the uncompressed image would be better than those obtained in the study for FCN. Additionally, there is no information on the type of compression used.
3) Testing the method:
3.1) The algorithm for determining the numerical value of VB for the proposed method has been quite briefly described. It is puzzling how the port algorithm treats the area of the photo with a gap in the edge (center of photo 12d).
3.2) Were only 8 edge images used for the method comparison (Figure 14)? If so, is it justified to determine the average error rate for such a small number of tested samples? If more was used, it might be worthwhile to provide a distribution of the results.
3.3) If the authors compare the FCN method to the K-means and Qtsu threshold, it would also be worthwhile for these methods to provide a numerical value for the flank maximum wear value VB and put this together on a graph (as in Figure 14)
Specific comments:
1. The names of the authors of the cited works are incorrectly cited in the text (lines: 49, 67, 69, 74, 76).
2. Figure 1 (the experimental setup) is not very clear, maybe the authors would consider using larger photos, additionally it seems that the order of photo captions "machine tool" and "machine center" is incorrect.
3) The authors use expressions and abbreviations that are crucial from the point of view of understanding the course of the research, but are not explained in the text of the article:
a) formula 1 - the right side of the pattern is a bit unclear
b) what is the "ideal image" and how was it generated? (line167)
c) furthermore the expressions: Adam (289), LebelMe (263), ReLu (212) are not explained at all and it is not described why these tools were used and with what parameters.
d) Table 2 essentially duplicates the information in Figure 7.
Author Response
Please see the attachment.
Author Response File: Author Response.docx