Dual-Kernel-Based Aggregated Residual Network for Surface Defect Inspection in Injection Molding Processes
Round 1
Reviewer 1 Report
The paper is novel, interesting, and is very well organized.
The work is original, the literature is well covered, the models developed thorough exposed and explained.
Moreover, the model developed is applied to an industrial scenario and the results obtained are well documented and analyzed.
In my opinion the data presented seems reliable and it is a study worth publishing.
There are a few grammar misconceptions throughout the paper and the authors could have avoided using the first person throughout the document, but apart from that it is a pretty neat document.
Author Response
The authors appreciate the reviewer for valuable comments. According to the comments, we sincerely revised our manuscript. Please refer to the attached file for response to the reviewer in details.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Low novelty. The authors may want to argue that the structure in Figs. 5 and 6 is new, but such an approach is not unusual. If the novelty of this paper lies in other aspects, that novelty was not shared with the reviewers, so I think additional explanations are needed.
- There exist lots of similar work for crack detection of concrete. Why don't you add these literature? including
- Dung, C. V. (2019). Autonomous concrete crack detection using deep fully convolutional neural network. Automation in Construction, 99, 52-58.
- Deng, J., Lu, Y., & Lee, V. C. S. (2020). Concrete crack detection with handwriting script interferences using faster region‐based convolutional neural network. Computer‐Aided Civil and Infrastructure Engineering, 35(4), 373-388.
- Chun, P. J., Izumi, S., & Yamane, T. (2020). Automatic detection method of cracks from concrete surface imagery using two‐step light gradient boosting machine. Computer‐Aided Civil and Infrastructure Engineering.
- Yamane, T., & Chun, P. J. (2020). Crack Detection from a Concrete Surface Image Based on Semantic Segmentation Using Deep Learning. Journal of Advanced Concrete Technology, 18(9), 493-504.
- The details of the study are unclear. I think it's necessary to know the computation time and how the training loss and validation loss decrease.
- I don't understand the meaning of the shaded area in Figs. 9 to 11.
- I think that you should give examples of images that failed to be classified and explain why CNN failed to classify.
Author Response
The authors appreciate the reviewer for valuable comments. According to the comments, we sincerely revised our manuscript. Please refer to the attached file for response to the reviewer in details.
Author Response File: Author Response.pdf
Reviewer 3 Report
This manuscript deals with the automation of quality inspection in manufacturing processes.
Convolutional Neural Networks are a set of an interesting approach for classification, and additionally used for anomaly detection in images and videos.
The authors deal with the inspection during an injection molding process (which is something new and not really investigated in the past(due to technical limitations regarding the process of images acquisition).
The study compares 2 CNNs implementations and demonstrated the feasibility of the use of theses Machine Learnings methods for enhancing Industrial Optical Inspection
The manuscript is clearly drafted and presented.
In the conclusion, it could be nevertheless, good to discuss the potential to use these approaches in combination with another one (sound patterns analysis, for example).
Thanks to the authors for showing another and challenging application of Machine Learning and particularly CNNs.
Author Response
The authors appreciate the reviewer for valuable comments. According to the comments, we sincerely revised our manuscript. Please refer to the attached file for response to the reviewer in details.
Author Response File: Author Response.pdf
Reviewer 4 Report
Authors present a dual-kernel-based residual network for the inspection of defects on surface during the manufacture of molding products through injection.
The analysis is clear and accessible. The work is interesting as it has direct potential applications in the manufacturing industry on the quality control. Authors explain the current drawbacks of the proposed approach before being applicable at the industrial process level.
I have minor comments and questions which should be addressed before publication of the manuscript in Applied Sciences.
) The introduction is quite extended compared to the total length of the manuscript. It is understandable for the authors to provide a literature background, but it is advisable to do it a in more concise way.
) Some figures appearing in the introduction are taken from other works. For example, Fig. 2 is Fig. 2 of Ref. 28. Have the authors obtained the corresponding copyright permissions? Also, is it indispensable to have such schemes and figures in the present manuscript? I do not think so, so they can be removed without loss of information. Including published figures is more common to review articles than original contributions.
) As experiments were conducted multiple times (for example 10), authors could include in the corresponding figures not only the average quantities but also the related error bars.
) On the classification performance authors state that it is slightly lower than the DAGM data but has the potential. This is important for implementation that the manufacturing level so authors could wish to extend the discussion by suggesting potential improvements.
) Line 159: “The same … in parallel”. Sentence is unclear and should be rephrased.
Author Response
The authors appreciate the reviewer for valuable comments. According to the comments, we sincerely revised our manuscript. Please refer to the attached file for response to the reviewer in details.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
I think the appropriate corrections have been made.
