U-Net-Based Foreign Object Detection Method Using Effective Image Acquisition System: A Case of Almond and Green Onion Flake Food Process
Round 1
Reviewer 1 Report
General:
The article appears to be interdisciplinary and certain level of extent it has embodied it!
There are major concerns to improve the quality of the research.
Comments:
Title: ……. Foreign Object Detection Method Using Effective Image Acquisition System
- a) Does this work involve deep learning algorithm(s) based object recognition or Otsu algorithm based image detection (threscholding)?
Abstract:
Do you think the formulation of problem and collection of synthetic dataset proposed in the article is suitable? In your article the training dataset/test dataset written in Figure 1 like insects, wood debries, plants, paper scraps, metal part, plastic scraps are suitable. However, during the Almond food manufacturing process (refer below there are few elements consider as dataset) Or Almond Harvesting?
In Almond food manufacturing process,
The almonds are delivered to the processing facility and are dumped into a receiving pit. The almonds are transported by screw conveyors and bucket elevators to a series of vibrating screens. The screens selectively remove orchard debris, including leaves, soil, and pebbles. A destoner removes stones, dirt clods, and other larger debris. A detwigger removes twigs and small sticks. The air streams from the various screens, destoners, and detwiggers are ducted to cyclones or fabric filters for particulate matter removal. The recovered soil and fine debris, such as leaves and grass, are disposed of by spreading on surrounding farmland. The recovered twigs may be chipped and used as fuel for co-generation plants. The precleaned almonds are transferred from the preclear area by another series of conveyors and elevators to storage bins to await further processing..
In the GOF, Onion, powder manufacturing process?
“Synthetic method to efficiently acquire training dataset of deep learning that can be used for food quality evaluation and food manufacturing processes”
The article argued effective acquire training dataset through deep learning. However, there is no specific deep learning architecture discussed in the article. However, U-Net is more suitable to detect microscopic images. Additionally, the researchers did not indicate neural network for segmentation of images mixed with foreign objects (e.g. in this case, Almond food process or Almond manufacturing process and at GOF)
It is not clear, the foreign object detection through Otsu algorithm or DNN model. Or Otsu for DNN.
If the result focuses on F1 score based on detection of foreign object, if the DNN model is used what is the accuracy on object/image detection in finding foreign objects. Food 101 is a collect of various end product of food ( a burger or Almond butter cake) Does your result significantly find the almond involved in the type of food? Is there any reason that you include Food -101 dataset with Synthetic dataset?
Section 2.5 must have the research result rather explaining precision and recall results
Introduction
Article focused on Food Manufacturing process, or Food processing or Almond processing. Or Almond Harvesting?
The reason for my question, is, I am not able to see the significance of the synthetic dataset used in this research!
Author(s) must understand what Almond Harvesting, Almond Food manufacturing process, is what the research gap to conduct this research is. Author(s) must understand a clear knowledge in food manufacturing and harvesting of each food items. The article use image acquisition using an electronic setting/table. There are good techniques available in DNN model to deal with image understanding.
Materials and Methods:
Figure 3 is a foreign object detection using Otsu algorithm In 2.3.4 the author claims to DNN object detection requirements of RMF, training and testing images. In image-based machine learning tasks, there a) image classification, b) object detection c) image segmentation. If you target these three-in-one solutions, look at its challenges and objectives.
Results:
Line 297 to evaluate the performance model learns the features” which performance model? DNN or Otsu method?
Line 306 refer the section 2.3.5, Section 2.3.5 uses the U-Net common architecture. How does it fits to your dataset? How many layers. In line 203, “architecture can learn the complex structures” what do you mean as complex structure? Where did you define it? If there is no architecture diagram of segmentation, for Almond images, and GOF images mixed with foreign objects would be useful to understand. Author has to add an example of max pooling. If the authors used U-Net how the detection of foreign objects has to be in a diagram.
Table 3 use DLFOD, HBFOD, it does not discussed in either in Abstract, Introduction or in Section 2.3.5!!! What are these results mean values signifies!? F1 Score or Mean?
Conclusion:
Author(s) made a conclusion of cost and time-consuming process… However, there is no time complexity graph or training graph mentioned or depicted. It looks like more than one author had written the conclusion based on their experiments!! What is the achievement of this article? Generation of images for deep learning image segmentation? Line 400, 401, 402 , 403 does it related to this research?? or future works?
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The work is about detecting foreign objects in raw materials for food, including almonds and green onion flakes. The work is interesting and well-presented. The experiments are also quite good.
A major issue is the novelty of the work. The data synthesis technique, the DNN model etc. are all existing work for example, U-Net is used and it is just something very typically used for segmentation in images. The overall process is novel and intuitive, but the technologies used are all existing ones. Thus, the novelty of the proposed work is quite limited.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The authors has substantially revised the article with additional paragraphs, figures to make the reader more interesting to go through the article.
1.
Figure 1 (c) moving objects detection, Are you using video extract frame images for the training? Where is the methodology for this?
2.
Figure 2 What is the purpose of showing the labelme tool? Are you plan to replace this with what type of deep learning algorithm?
“Our proposed method, likewise the color sorting machines, focuses on the RMF that can be easily obtained in food inspection. However, it uses DNN to consider not only color but also various features such as shape, texture, and size. To train the normality of RMF, several images of RMF mixed with various objects are required”
Does your problem focused on multi-class labelling classification? I wonder the object detection works well only on GOF objects as you claimed Korea has this debris found during almond food processing.
3.
From your response 3, your reply on U-Net not used in RMF and there is no significant references. Does it mean, your foreign object detection able to detect micron objects mixed with almond food manufacturing process? Which debris in your table has the micron units? I really wonder why not you are able to compare with recent advancement of DNN object detection algorithms or you would have use to efficient parameter tuning to achieve this… It would be good to refer recent DNN object detection algorithms and compare with your results and discussion.
- Results and discussion: U-Net with other DNN works? Hard to measure the performance of U-NET as there is no significant comparison. Reference 37 using U-NET/ CNN pre trained model has better results. You work has adapted U-NET significantly, It would be good to change the Title
U-NET based FODM for Effective image acquisition: A Case of Almond Food Process.
ADD U-NET ORIGINAL ARTICLE AS YOUR REFERENCE.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
