AI-Driven Crack Detection for Remanufacturing Cylinder Heads Using Deep Learning and Engineering-Informed Data Augmentation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Research oriented towards practical application demands holds significant value in both fundamental and applied research. An efficient machine vision-based method for detecting cracks in components is realized through sample enhancement techniques.

However, judging from the test results and considering the needs of engineering practices, the current accuracy rate is relatively low.

As an enhanced research approach, in the "Crack and Porosity Detecting Set," metal crack detection agents, such as solvent-based dye penetrant DPT, are commonly used to significantly enhance the visualization of cracks. This method is also recommended by standards such as DIN 54152 and MIL-I-25135 D. It is suggested that the authors adopt similar methods for sample pretreatment in their subsequent research to enhance the samples from the source.

Author Response

Dear Reviewer, Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

For accurate crack detection, this paper combines the existing crack detection model with a synthetic data generation technique. The paper is well-organized and written. However, there are some limitations.

1. This paper was named AI-Driven Crack Detection, but the main focus is on traditional data generation, which has been outperformed by other AI-driven techniques, e.g., GAN. Why do you use synthetic data generation techniques.

2. In the data generation process, this paper extracts different cracks and integrates them with collected data. How to prove the feasibility of this method.

3. For the CNN-based detection model, the function of synthetic data generation to the detection model cannot be clearly identified. How to determine the integration method, including the location, type and size.

4. During the data generation process, have you ever considered the similarity of the training set, which would affect the training performance.

5. The novelty and contribution of this paper is limited.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

Dear Reviewer, Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposed a data augmentation method for crack detection of mechanical components by randomly copying the cracks into other places in the images. I have two major concerns about this paper:

1) the novelty of this paper: the operation of randomly copying one crack to another position are naturally inspired by the commonly-used data augmentation methods in computer vision, i.e., randomly rotating or cropping the training images. Could the authors give more specific explanation about why this method is novel for crack detection? e.g., what is the reason why it improved the performance of object detection model on this task? Or what kind of challenges it conquered in this specific industrial application?

2) The soundness of this paper: The experiments of this paper argued that recall is more important than MAP. I cannot agree with this opinion. It is a biased evaluation that the results only hightlight the improvement of recall but neglect the precision of the result. I think the increase of recall is related to the operation of copying cracks in training data images, that is, more randomly copied cracks result in more false alarms. I would suggest the authors clarifying the negative impacts caused by the data augmentation method.

Comments on the Quality of English Language

This paper doesn't have major language problems.

Author Response

Dear Reviewer, Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The paper is well modified and can be accepted in present form.