Real-Time ConvNext-Based U-Net with Feature Infusion for Egg Microcrack Detection
, Yuejia Li, Xin Jiang, Wenxin Sun, Changsheng Zhu, Yuanzheng Mo, Shaojia Yan and Chun Zhang *
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript presents a Real-time ConvNext-Based U-Net model with Feature Infusion (CBU-FI Net) for egg microcrack detection. I think it is worth publication after making revisions for the following comments:
Q1. Please explain the reason to use feature fusion in decoder part to improving segmentation accuracy.
Q2.In introduction part, it is difficult to follow the content. I suggest the authors to optimize the content about egg microcrack detection technology based on an electrical characteristics model and acoustic techniques. On the other hand, author should reorganize the content about deep learning based machine vision technologies for egg crack detection.
Q3. In this study, the author chose segmentation task instead the object detection task. In practice, it is meaningful to detect the egg crack by providing bounding box of crack. By employing the detection model, it is easier to solve the issues about the model’s scale and efficiency. Therefore, please give reasonable explanation.
Q4. For testing the proposed model, the author also use the publicly benchmarked
CrackSeg9k dataset. Firstly, did this dataset utilize for training the model? Secondly, images of the crack of road is obviously different to the images of egg crack. I suggest the author should replace the content about this with egg crack test conditions.
Q5. For the evaluation metrics, the detail description about the used metrics should be adding by providing the related equations.
Comments on the Quality of English Languageno comments
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe publication is part of the area of ​​scientific research supporting the process of identifying graphic information necessary for the qualitative classification of stored agri-food products using eggs as an example. The work uses and modifies known artificial intelligence (AI) procedures, in particular deep learning techniques of convolutional neural networks and methods of neural analysis of digital images. The authors of the work focused on improving methods of automatic detection of microcracks (constituting a class of so-called small identification targets) occurring in stored eggs. For real-time detection of microcracks in the structure of eggs, the U-Net model was proposed, which is a convolutional neural network dedicated to segmentation of biomedical images. The idea of ​​the proposed neural classifier is based on the well-known ConvNext concept, which includes the process of infusion of representative features (CBU-FI Net). For this purpose, a standard edge detection strategy (known from classical image analysis methods) was used, also taking into account the spatial continuity of cracks (presented in digital images). The authors proposed an original strategy for aggregating multi-scale features implemented in the decoder. This allowed for improving the extraction process of local graphic artifacts and improving the quality of the process of identifying global semantic information.
The structure of the convolutional classification network was modified, among others, by reducing the number of model parameters (to one third of the original U-Net), which significantly minimized hardware requirements. According to the authors, the proposed procedure optimizes the efficiency of the identification neural model and increases its efficiency.
The described identification method can be dedicated, among others, to supporting processes occurring during the transport and storage of eggs. The proposed algorithm for visual detection of graphic artifacts (in the form of cracks in egg shells) can also be a significant support in the process of qualitative assessment of selected agri-food products.
In terms of methodology and formality, the work is constructed correctly. It also contributes to the development of scientific knowledge in the field of applied sciences. In my opinion, such publications are needed because they have, among other things, a desirable utilitarian meaning. The reviewed work falls within the scientific area undertaken by the Agriculture journal.
However, the article requires several explanations and minor corrections, in particular of a stylistic and formal nature.
¾ the definitions of quantities used in formulas and tables should be corrected and supplemented; for the reader's convenience, it is best to place them directly under the formulas and tables,
¾ the full names of the acronyms used should be supplemented,
¾ it may be worth comparing the proposed procedure with other neural identification techniques, e.g. with the latest YOLOv9 algorithm (higher accuracy with fewer resources), which constitutes significant progress in the process of object detection,
¾ the conclusion lacks utilitarian remarks: e.g. areas of application, directions and prospects for the development of the proposed method,
¾ in the "References" chapter, the small share of citations of similar works published in Agriculture is surprising.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsLine 21 needs a reference.
In line 60-61, the authors are claiming that both the acoustic and electrical detection methods are destructive and invasive. However, they did not mention that earlier where they cited electrical detection method. This have to be mentioned earlier.
Reference is needed for CrackSeg9k (line 295-296).
What is the distribution for training-testing among the data?
Section 3.1.1 discussed some results of some methods. The authors mentioned some methods performed poorly and some methods improved performance. However, they did not explain why the poor performance occurred and by which change they achieved better performance. The reason must be stated.
Same comment as above for sections 3.1.2 and 3.1.3.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThis research is about Feature Infusion for egg microcrack detection. It is very interesting data that the egg consumption about 34.5 million tons in 2023. Especially focused on developing high-speed, high-precision poultry egg crack detection technologies.
The paper organization is very good. Unfortunately the visualization of some figures sizes are not proportional. For example Figure 2. The sub caption in images such Original Image and so on, the font size is bigger than the caption. Also in Figure 4 and Figure 6.
In line 273, "Where, α denotes the adjustment factor." should be "where, α denotes the adjustment factor." without the new paragraph.
In Figure 1, the arrow beside Basic Block should be in the proper position.
The caption on Figure 1: "Architectural design of the CBU-FI Net. The network consists of three levels of encoders and decoders." It's better it is mentioned or more explanation in the caption what is the three levels of encoders and decoders. Of course there is more explanation in the paragraph.
In Conclusion:
"Experimental results on a dataset comprising over 3,400 graded egg microcrack image patches demonstrate that CBU-FI Net achieves a parameter reduction to one-third of the original U-Net, with an inference speed of 21 milliseconds per image (1 million pixels)."
How did these 3,400 graded egg microcrack image patches represent the all consumption egg cases?
The quality of eggs does not only depend on the detection of cracks in the egg but also on the substances contained in the egg. It should be also mentioned in Introduction.
The similarity paper is good.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsNo more comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsComment 3 has not been addressed. The authors did not compare their method with any of the YOLO models. Please run YOLOv8 or later model on your dataset and tell us the result!
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
