A Size-Grading Method of Antler Mushrooms Using YOLOv5 and PSPNet
Round 1
Reviewer 1 Report
The paper proposes a model that integrates a single-stage target
detection of YOLOv5 and a semantic segmentation of PSPNet, and proposes a Y-PNet model for real-time segmentation. In addition, the authors use MobileNetV3 and the CBAM module with the aim of reducing the size of the model and improving other characteristics. The purpose of the model is the classification and segmentation of mushrooms in the production process.
It is particularly interesting that the authors implemented the model in a device that should classify mushrooms by quality level in factories.
Personally, I prefer when the program code is available with the papers, and a few pictures on which it can be tested, although in this case it is understandable that the authors made the program code unavailable. However, upon reasoned request, they would make the program code available.
On the other hand, they described the model clearly enough in the paper, so that it should not be a problem for interested people to implement it.
The paper does not revolutionize the field of application of segmentation algorithms, but it is of sufficient quality that it will be of interest to experts dealing with the classification and segmentation of agricultural products.
Cited references are relevant and authors have not self-cited.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper offers a novel approach based on YOLOv5 and CNNs to classify antler mushroom quality.
Many citations can be improved throughout the paper. In line 29, for example, papers 3 and 4 cited feel arbitrary, as they are not related to the research, are not the best papers in their themes and are not very recent works. It would be a better opportunity to cite a more comprehensive survey in visual applications of machine learning and/or applications in agriculture more closely related to the theme of the paper.
In lines 32 to 63, references 6 through 15 tell a story of the revolving timeline of mushroom detection and classification, from 1991 to the present day. The chronological order should be made clear in the text in order to justify referencing older papers.
The paragraph in line 64 restates a lot of information from the first paragraph and should be shortened.
The section on YOLOv5 (line 98) lacks any references. There are the seminal YOLO papers and several recent surveys on the subject that could reference the information provided.
The same issue occurs on the section on PSPNet (line 137), it has no references.
The text in Figure 1 is perhaps too small (smaller than the font size in the body of the text), and the Figure touches the paper's margin. It could be rearranged to solve both problems. Labels in the Figure (such as "slice" or "CBL") are also not explained. They should either be omitted from the Figure or explained in the text.
Figure 3 presents the proposed fusion of YOLOv5 and PSPNet from Figures 1 and 2. It should highlight the proposed interaction between YOLOv5 and PSPNet and the proposed improvements mentioned in lines 176 and 181.
I believe line 238 should say "time-saving" instead of "time-consuming".
Table 1's caption incorrectly mentions "experimental results", but should refer to dataset description. The "Label category" column name is also misplaced.
Equations 5, 6 and 7, (line 352 onwards) as well as the explanation on true/false positive/negative could be relegated to a reference to a book.
Line 381 mentions MnasNet is used for comparison but the results for it are not shown. It would be an important validation to the proposed method to use a different CNN from ResNet as comparison. MnasNet has a mobile focus, but there are other possible CNN architectures like (e.g. VGG).
In Table 2, I would suggest best results in each column/criteria be highlighted in bold for ease of comparison.
Starting from line 444, the term "size" classification is sometimes used instead of "quality" classification. Whether these terms are synonyms or not, this should be clarified throughout the text.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
1. Please, try to use short sentences, instead of complex ones.
2. the sentences from line 64 to 66 should be divided and placed into line 27
3. line 66-68 : please clarify that you mean instant segmentation
4. line 68: conventional term object detection instead of target detection!
5. lines 69-70: please, do no use passive voice, if direct form acceptable.
6. lines 72-83: this should be placed in another section;
6. lines 72-83: please, compare your solution with yolact++ and other architectures for the real time instant segm which you may find, for instance, here: https://paperswithcode.com/task/real-time-instance-segmentation
7. lines 72-83: MBNet v3 and CBAM need to have references
8. Just in case, today YOLO5 provide possibility to instant segmentation.
9. Just in case, today beside yolo5 you may use yolo6, 7, yolor and other modification of this approach.
10. lines 72-83: Please arguing selection of yolo5 and pspnet and mbnetv3
moreover, mbnetv3 was designed for Google pixlephone, i'm wondering if you use it as base for network implementation.
11. lines 107:109:basic yolo5 architecture is mainly copy of yolo4, given in pytorch framework.
12. line 111: we use augmentation instead of enhancement
13. fig 1: please ,exceed the figure size
14. fig 3: please ,exceed the figure size
15. If you have the task of only one object segmentation than why do you need object detection, you may just use semantic segmentation!
16. line 224: resnet does not enlarge receptive filed, skip connection introduce diversity into receptive field
17: lines 237-253: please, for the first show MBNet V3 architecture you are using.
18. Please, describe how do you use netadapt.
19. h-swish (x), is well known function, wht do you call it new?
20. SE) module does not reduce as it slef, only by pointwise conv.
21. Please take references for all modules you are talking about.
22. I could not any CBAM in Fig. 3, wher dou you use it, and how do you combine it with SE layer
23. line 273-275: the statement is no correct: "Because CBAM is a widely used lightweight attention module, 273 it can be easily embedded into the basic CNN for end-to-end training, and the module takes up few parameters and computations after embedding, which can be neglected"
24. table 1: please, provide the examples of images and their labeling
25. It is better if conclusion contains numerical results, their novelity and planes of its improvement.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
1. Introduction: I'm doubting of using new category of CV tasks : real-time segmentation algorithm, or you need to take some definition, for instance, object localization (OR OBJ.DET + full image segmentation). In any case, it would be better to categorize you task as some conventional type.
2. Point 7：lines 72-83: please, compare your solution with yolact++ and other architectures for the real time instant segm which you may find, for instance, here: https://paperswithcode.com/task/real-time-instance-segmentation% Here I also meant that you when you are talking about any conventional tasks you need to say about state of the art here and justify why do you need another solution.
Point 11：lines 72-83: Please arguing selection of yolo5 and pspnet and mbnet 3 moreover, mbnetv3 was designed for Google pixlephone, I'm wondering if you use it as base for network implementation - I think you do not answer and make change in accordance to this remark
fig1 and 5 are now more readable, but I still recommend you to increase their height
Point 23 and fig 5: the position of CBAN is not intuitively understandable
Point 25：It would be good tone if you show examples of database you are using while describing dataset
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
