An Improved YOLOv5 Algorithm for Bamboo Strip Defect Detection Based on the Ghost Module
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsReviewer's comments on the paper titled “Improvement of YOLO detection strategy for defects in bamboo strips” for the MDPI Journal FORESTS
The current paper is an original contribution, which corresponds to the scope of the MDPI Journal Forests. A lightweight YOLOv5s neural network algorithm with the using of a Ghost module was suggested to identify the five main surface defects of bamboo strips. It was stated, that the efficiency of the YOLOv5s neural network algorithm application can be increased by the introducing an attention mechanism CA module. It enables to improve the recognition ability of the model target. The research also implements a C2f model to enhance the network performance and the surface quality of bamboo strips. Advantages of the lightweight YOLOv5s neural network algorithm with the using of a Ghost module were tested by the experiment. The main advantages are improved detection speed and increasing the recognition accuracy.
The obtained results are adequately presented and the paper is enough clear for understanding. But at the same moments some comments should be addressed to made more clear some technical details. The comments are:
1) Title of the paper, probably, can be expanded to made it more clear. It can be “Improvement of YOLO detection strategy for defects in bamboo strips using the Ghost module”.
2) Chapter 1 “Introduction” can be expancec by the information regarding the the practical applications of the bamboo strips. This information enables to understand in the more details significance of the defects detection in the bamboo strips.
3) Chapter 4 “Experiment results and discussions” can be completed by the some figures, preferably photos, which can illustrate the experiment, which was carried out.
4) There are no clear visible differences between the Figure 1 and Figure 9. Contents of the figures is practically the same, so as the same specimens were shown.
5) Conclusions of the paper are too short and needs to be expanded by the information regarding the results, obtained in course of the current study. The major numerical results should be added also so as enables to understand the main statements were done.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors The manuscript discusses new possibilities for detecting defects on the surface of bamboo strips. The research topic is in line with the environmental trend of replacing plastic with bamboo. This work uses computer vision methods and neural network algorithms for automatic defect detection. The structure of the manuscript corresponds well to the logic of the research. To select the most effective research path, the authors analyzed possible research tools such as YOLOv3, YOLOv4, YOLOv5, and YOLOv9. The improved structure of YOLOv5s is proposed and implemented in the manuscript (lines 209, 210). New modules are developed (Lines 222, 223, 257, 258). Comparative experiments (Section 4) showed that the practical application of the new research results provides an optimal balance between the amount of calculations and the accuracy of the results, which impart a competitive advantage in detecting bamboo strips with various types of defects (Lines 401-428). The text is professional; Figures 1-9 and Tables 1-5 enhance understanding of the methodology and results of the study. The statements and conclusions are logical and supported by relevant references. Thus, the manuscript contains new scientific results in the field of automatic detection of defects in bamboo strips, which are important for practice and will be of interest to many potential readers of Forests. In addition to contributing to the development of technologies for automatic detection of defects in plant-based materials, the manuscript may inspire Forests readers to search for new ways to grow bamboo with a minimum number of defects. However, this will not diminish the importance of automatic detection of these defects. Comment: Should formula (3, Line 311) contain a 100% multiplier? Please check, taking into account formulas (1) and (2).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors presented a new version of YOLO (lightweight YOLOv5s) to detect defects in bamboo strips. I have some critical concerns that should be addressed.
1- Please explain why we should use the deep learning models for the image analysis with citations. There are good references that clarify deep learning methods overperformed the traditional ones. For example, the authors can use the following sentences with citations that clarify this matter.
Traditional image analysis techniques, which use basic image features for object segmentation and detection, often miss complex spatial relationships. Deep learning models, however, excel in automatically extracting high-level semantic features, offering more robust solutions in this area【x】【xx】.
【x】 https://doi.org/10.1109/ACCESS.2024.3385425
【xx】https://doi.org/10.3390/app11209691
2—The literature review is not as complete as this. Please provide more literature on this; if it is not cited, please cite it and give us more Gaps in this area. For example,
Comprehensive defect detection of bamboo strips with new feature extraction machine vision methods
http://www.jamstjournal.com/en/article/doi/10.51393/j.jamst.2023018
3—In Figures 8 and 7, please provide the loss curve of training and Val loss and mAP of training and validation. This can help us improve the network's performance during training.
4—The Yolo results are more of a bonding box, but the authors don't show the visual results to show the false and misdetected areas. Please provide 6 subfigures for each defect and show the performance of the proposed method.
5- Please provide the limitations of the presented work.
6- Please provide the computation time for each considered method and the proposed method for comparison (training time per epoch).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsLine 10: "gmaill" - is this correct?
Line 177: Chapter 3 is only about Methods (of detection), not Materials (Bamboo strips), chapter title should be modified.
Line 280: Add a reference to Fig. 1 here
Line 285: 640x640 pixels are squared-shaped images, but the images in Fig. 1 are rectangular?
Line 287: That means that approximately 4800 images were free of defects?
Line 305: Does the specification "FPS" in the following tables mean that so many images per second supplied by a camera can be analysed? That would be an important indication of the performance of a real detection system - should be included in "5. Conclusion"
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsI dont have further comments.