An Algorithm for Real-Time Aluminum Profile Surface Defects Detection Based on Lightweight Network Structure
Round 1
Reviewer 1 Report
In this work, the authors investigated a lightweight network based on the YOLOv5s algorithm to realize a real-time defect detection of aluminum profile surface on the embedded system with promising detection accuracy. The manuscript is rather well prepared; however, the following comments can help improve the manuscript:
· The novelty of the work is not clearly highlighted.
· What is the highlight result by the research?
· What is the industrial application of your research?
· Figures need to have appropriate scale bars.
· The conclusion section needs to provide more detailed information, quantitative rather than qualitative.
Author Response
Dear Review,
Thank you for your suggestions for this article, starting with a point-by-point response to the comments.
Comment 1: The novelty of the work is not clearly highlighted.
Author response: We really appreciate for your valuable advice. We proposed an efficient and lightweight defects detection method for aluminum profile to realize a real-time surface defects detection on the premise of guaranteeing detection accuracy. We embed the attention mechanism into the ghost module to form the backbone layer of our lightweight network. The size of the model is reduced while the accuracy of the algorithm is guaranteed. And the depthwise separable convolution replacing the regular convolution in Neck significantly reduces the number of model parameters and speeds up detection speed.
Author action: To emphasize the novelty of our work, we have rewritten the specific innovations of this paper. (Lines 121-131)
Comment 2: What is the highlight result by the research?
Author response: We propose the lightweight network model based on YOLOv5s. The main advantages of the new algorithm are real-time detection and high accuracy. And the algorithm can work in embedded devices.
Author action: We further highlight the highlights of the study in our conclusion. (Lines 355-365)
Comment 3: What is the industrial application of your research?
Author response: At present, the volume of surface testing instruments in the industry is too large, our instruments are small and easy to carry. It is mainly convenient for sampling inspection of quality inspection departments in industrial production.
Comment 4: Figures need to have appropriate scale bars.
Author response: We really appreciate your valuable advice.In image analysis, we usually analyze the pixels of the image. Because there is uncertainty in detecting the size of the object, the focal length is usually a dynamic quantity, resulting in inconsistent image scale. To solve this problem, most articles explain the pixels of the overall dataset with annotated pixels. The ratio of annotated pixels to image size is used to determine the type of detected objects. A mere 1.23% of the annotated pixels belong to small objects. For example, the average pixel size of the pinhole in this article is 20×20, and the image size is 640×480. annotated pixels is 0.13%, so it is a small target type.
Author action: We added pixel descriptions of the overall dataset, the introduction of the four defects and pixel descriptions. (Lines 137-143)
Comment 5 : The conclusion section needs to provide more detailed information, quantitative rather than qualitative.
Author response: We gratefully appreciate for your valuable suggestion.
Author action: More detailed information is provided in the conclusion section. (Lines 355-365)
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Title: An Algorithm For Real-time Aluminum Profile Surface Defects Detection Based On Lightweight Network Structure
Authors: JUNLONG TANG, SHENBO LIU, DONGXUE ZHAO, LIJUN TANG, WANGHUI ZOU, BIN ZHENG
The manuscript by J. Tang et al. deals with the algorithm of the novel lightweight network for the real-time detection of aluminum profile surface defects. The authors propose the lightweight network model based on YOLOv5s. The main advantages of the new algorithm are real-time detection and high accuracy. What is important is that the test is performed for 4 types of defects with confidence score above 0.8. In my opinion, this paper can be recommended for publication.
Two points require clarification:
- - It would be useful if the authors added a short description of major types of surface defects of aluminum profile. What about non-conductive defects? Or about pick-ups?
- - The four types of aluminum profile defects are shown in Figure 2 (pinholes, soiling, folds, scuffing). If the authors believe that these are typical defects on the surface of aluminum profile (page 3, lines 130-131), why there is a discrepancy with the four types in testing the algorithm accuracy (pinholes, scratches, dirty spots, folds). I think it is more correct to illustrate exactly the types of defects being detected in the test (Section 4.3, page 9).
Author Response
Dear Review,
Thank you for your suggestions for this article, starting with a point-by-point response to the comments.
Comment 1: It would be useful if the authors added a short description of major types of surface defects of aluminum profile. What about non-conductive defects? Or about pick-ups?
Author response: We gratefully appreciate for your valuable suggestion. Non-conductive defects usually refer to internal defects. Our study is on surface defects. We detail the types of defects according to the definition of literature [34] and the causes of defects.
Author action: We have added the description of 4 types of defects on the surface of aluminum profiles. (Lines 137-143)
Comment 2: The four types of aluminum profile defects are shown in Figure 2 (pinholes, soiling, folds, scuffing). If the authors believe that these are typical defects on the surface of aluminum profile (page 3, lines 130-131), why there is a discrepancy with the four types in testing the algorithm accuracy (pinholes, scratches, dirty spots, folds). I think it is more correct to illustrate exactly the types of defects being detected in the test (Section 4.3, page 9).
Author response: Thank you very much for your help. We are sorry, due to our mistake. We were not able to achieve uniformity in the names of the defects. The four types of defects are Pinhole, Scratch, Dirt, Fold.
Author action: Already changed. (Lines 137-143)
Author Response File: Author Response.pdf
Reviewer 3 Report
The work is very interesting and contains important issues regarding the detection of defects. This article is suitable for publication but requires minor changes. Please note the following notes:
1. Were tests performed on materials other than aluminum alloys?
2. Have other non-destructive methods such as C-scan or recurrence methods been tried to detect defects?
3. Please cite other works on defect detection because there are only 37 items in the bibliography? I suggest to refer to works on recurrence methods and C-scans because these works are also related to non-destructive methodology of defect detection. In recurrence methods, algorithms are also used to detect defects.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
