Improved YOLOv5 Network for Detection of Peach Blossom Quantity

Round 1

Reviewer 1 Report

The database used in the study was sourced solely from the peach tree breeding base in the Mancheng District Branch of the Natural Resources and Planning Bureau of Baoding, focusing on three-year-old varieties like "Jiuli", "Chunxue", and "Baojiajun". This limited scope raises concerns about the algorithm's performance across diverse settings. The research did not verify the algorithm's effectiveness on images from other observation locations, potentially affecting the model's generalizability and applicability in different agricultural environments.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In the introduction section, the motivation for Improved YOLOv5 Network for Detection of Peach Blossom Quantity is week

I tried to find the literature review section but couldn't. There is no literature gap. How does the author claim the novelty?

What is the author's contribution in Figure 3. Structure of CAM ?

It would be better how the equations (From 1 to 12 ) are helpful for Improved YOLOv5 Network for the Detection of Peach Blossom Quantity  . In each case author suggested giving an example and showcasing the improved YOLOv5 Network for the Detection of Peach Blossom Quantity.

I tried to find the computational requirements and processing times associated with deploying the improved YOLOv5 model for peach blossom quantity detection on different hardware platforms but nowhere the author mention it.

The author needs to compare the existing approaches to traditional methods for peach blossom quantity detection and see what impact the improved YOLOv5 architecture has on the accuracy, precision, and recall rates

The dataset description is available but their source is not available 

In terms of scalability, how does the improved YOLOv5 network perform as the size of the peach orchard or the quantity of blossoms to be detected increases?

Ok 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

accepted.

Author Response

Thank you again for your time and effort, and for your interest and support for our paper. 

Reviewer 2 Report

I am shocked the revised version of the manuscript was not highlighted. Whatever question was raised that is not highlighted In the manuscript? 

Authors are requested to provide a highlighted version once again 

I have gone through the details of the response, there also the author has not mentioned which section, line number, etc.

ok

Author Response

Dear Reviewer:

We are very grateful for the opportunity to have our paper, titled "Enhanced YOLOv5 Network for Instance Segmentation of Various Peach Blossom Morphologies", reviewed by you and for your valuable comments and suggestions. We have carefully studied all the comments and made the corresponding revisions, which are explained in detail in this response letter and marked in red.

Comment1: In the introduction section, the motivation for Improved YOLOv5 Network for Detection of Peach Blossom Quantity is week

Response: We acknowledge that our paper did not sufficiently explain our research motivation and significance in the introduction section. To improve this, we rewrote the introduction section with the following logical order: we briefly introduce the current situation of the peach industry (lines 21-26); we illustrate the importance of peach blossom quantity detection for peach tree growth and management (lines 27-40); we state that the main purpose of peach blossom number recognition is to assist thinning (lines 41-44); we present different thinning methods and their advantages and disadvantages, as well as the future direction of thinning technology (lines 44-55); we introduce the traditional image processing methods for fruit tree flower detection and their drawbacks (lines 56-70); we introduce the deep learning techniques for fruit tree flower detection and their limitations (lines 71-83); we add the YOLO series of fruit tree flower detection (lines 84-89); we point out the challenges and limitations of peach blossom detection in complex environments, highlight the contributions of our study, propose our improved YOLOv5 network, and outline its main features and benefits (lines 92-125).

Comment2:I tried to find the literature review section but couldn't. There is no literature gap. How does the author claim the novelty?

Response: You pointed out that you could not find the literature review section, which might be due to the unclear structure of our paper. We apologize for the inconvenience caused. We have revised the literature review section and report the following: We have made a comprehensive revision of the reference section and the literature review section of our paper: We have added some recent literature related to our research topic, to reflect the current research status and trends. These literature include [TIAN Y, 2019; Farjon, 2020; TIAN Y, 2020; Tao K, 2022; Ye X I A, 2023; SHANG Yuying, 2022, etc.]. We have adjusted the structure and content of the literature review section (from machine learning to deep learning and then to the specific application of the YOLO series in fruit tree flower recognition), to more clearly state the purpose and contribution of our research. We have added a paragraph at the end of the literature review section, summarizing the main findings and gaps of the literature review, as well as the contribution of our research. These are the revisions of the reference section and the literature review section of our paper. You can find these revisions in the introduction section of the revised manuscript we submitted (line numbers: 56-91).

Comment3:What is the author's contribution in Figure 3. Structure of CAM ?

Response: We have made corresponding revisions to address the issue you raised, and refined the specific operations, as follows: Since tiny bud detection requires contextual information, our first contribution in this part is to add a CAM module on top of the original Neck layer. The specific operation is: we perform dilated convolutions with different dilation rates on the 9th C3 layer to obtain contextual information with different receptive fields to enrich the FPN’s contextual information (at the same time, we also fine-tuned Figure 3: Figure 3. Overall network structure. CAM and FSM are the main components of the network. CAM injects contextual information into FPN, and FSM filters FPN conflicting information (line number: 170), and marked the FPN structure in the figure). The second contribution is to explore the fusion methods of CAM and FPN (Concat fusion and adaptive weighted fusion). The results show that CAM addition is effective, Concat fusion method improves 2% compared to the original YOLOv5, and adaptive weighted fusion method improves 3.1% compared to the original YOLOv5 (line number: 203-209).

Comment4:It would be better how the equations (From 1 to 12 ) are helpful for Improved YOLOv5 Network for the Detection of Peach Blossom Quantity  . In each case author suggested giving an example and showcasing the improved YOLOv5 Network for the Detection of Peach Blossom Quantity.

Response: According to the comments you gave, we have made corresponding revisions. Equations (1-7) show the feature refinement process of the FSM module to filter out conflicting information. To verify the effectiveness of the FSM module, we added Table 2: Comparison of FSM and Concat under YOLOv5 (line number: 250). The results show that replacing Concat with the FSM module can indeed improve the network accuracy, with a 4.8% difference from the previous mAP. This indicates that the fusion of feature maps at different levels does bring conflicting and redundant information, and the feature refinement by the FSM module can further suppress these conflicting information and improve the accuracy of peach blossom target detection. Equations (8-10) represent the K-Means++ custom anchor box; equations (11-12) represent the loss function, and the effectiveness of these two modules is also listed in Table 5. Comparison of models during the overall ablation test (line number: 371).

Comment5:I tried to find the computational requirements and processing times associated with deploying the improved YOLOv5 model for peach blossom quantity detection on different hardware platforms but nowhere the author mention it.

Response: We have made the following response to the question you raised: This paper uses FPS (frames per second) to indicate the speed of different models. FPS (frames per second) is a metric that measures the inference speed of the model, indicating how many images the model can process per second. (Line number: 360-362)

Comment6:The author needs to compare the existing approaches to traditional methods for peach blossom quantity detection and see what impact the improved YOLOv5 architecture has on the accuracy, precision, and recall rates

Response: According to your suggestion, we have added two evaluation metrics, precision and recall rates, to the analysis of the results of the comparison of the peach blossom recognition and detection methods of the mainstream models. A total of five evaluation metrics are used to analyze the recognition and detection performance of the improved YOLOv5 model compared to other recognition and detection models. The analysis of the P and R values shows that: SSD and Faster R-CNN have relatively high R values, but their P values are below 80%. Overall, the P and R values of SSD, Faster R-CNN, YOLOv3 and YOLOv4 are significantly lower than those of YOLOv5s and our improved YOLOv5s algorithm. As can be seen from Table 5, YOLOv5s achieved good results in all indicators (P: 83.3%; R: 90.4%; mAP: 83.1%), and the model inference speed was 45.61 frame/s, which is acceptable due to the high density of peach blossoms in a single image. The improved YOLOv5 algorithm achieved a P value of 90.1%, which is 6.8 percentage points higher than the original YOLOv5 algorithm; an R value of 93.2%, which is 2.8 percentage points higher than the original YOLOv5 algorithm; and an mAP value of 0.902, which is 7.1 percentage points higher than the original YOLOv5 algorithm. You can find the corresponding revisions in the “2.2. Performance Comparison of Mainstream Target Detection Models” section (line number: 393) of the revised manuscript we submitted.

Comment7:The dataset description is available but their source is not available 

Response: We have made the following response to the question you raised: Due to the small morphological differences of peach blossoms among different varieties and regions, the peach blossom acquisition area of this experiment was initially only selected from the peach tree breeding base of the Natural Resources and Planning Bureau of Mancheng District, Baoding City, Hebei Province. The reason why we chose the data from the peach tree breeding base of the Natural Resources and Planning Bureau of Mancheng District, Baoding City, Hebei Province as the experimental data is that the base is a national key peach tree good variety breeding base, with a large number of peach tree varieties, and professional botanists have annotated the different peach blossom morphologies, providing a reliable data source for our algorithm. Although the flower morphology of peach blossoms of different varieties and regions is similar or even the same, the cultivation methods and tree shapes of the orchards will have some impact on the quality of the peach blossom dataset, such as the occlusion of other peach trees or their own branches and leaves. Photos are two-dimensional images, and if the tree shape is complex, the occlusion of flowers and branches is very complex, it is impossible to accurately identify the number of peach blossoms, resulting in a large difference in the recognition performance of the model. The peach trees used to collect images in this study are modern orchard cultivation methods, with appropriate row spacing and standard tree shapes, which are Y-shaped and one-word-shaped, which can minimize the occlusion of flowers and branches, making the dataset quality higher. According to your comment on the limitation and unavailability of the dataset, due to seasonal factors, we cannot obtain more peach blossom sources at present, but we will also consider this aspect in the later stage, obtain more extensive data sources, and further improve the generalization ability of the model. We also mentioned the relevant content in the discussion section of the paper (line number: 460-462).

Comment8:In terms of scalability, how does the improved YOLOv5 network perform as the size of the peach orchard or the quantity of blossoms to be detected increases?

Response: The improved YOLOv5 network proposed in our paper first adds a dedicated output layer for small target detection on top of the original three output layers; then adds a combined context extraction module (CAM) and feature refinement module (FSM); and finally uses the K-means++ algorithm to cluster and count the range of multi-scale channel elements, and adopts a novel bounding box regression loss function (SIoU), to improve the detection accuracy and ensure efficiency. To verify the scalability of our proposed model, we conducted experiments on different numbers of peach blossom datasets (divided into small, medium, and large peach blossom scales, each scale selecting 100 images that meet the peach blossom number requirements), and the results are shown in the following table:

As can be seen from the table, our network can maintain a high detection accuracy and speed for different numbers of peach blossoms, among which the large-scale peach blossom detection accuracy and speed are relatively low, but still within the acceptable range, indicating that our network has good scalability. Our model also has obvious advantages compared to other mainstream models, as shown in the speed and accuracy comparison of Table 6. Comparison between different object detection models (line number: 396) of the revised manuscript, it can be seen that our model is faster than other models except for being slightly slower than the original YOLOv5, and has the highest detection accuracy. These results demonstrate the superiority of our network in terms of scalability. We hope these answers can satisfy the reviewer’s questions, and we are happy to discuss further if there are any other issues.

We thank the reviewers again for their comments and suggestions. The above are the revisions and explanations we made in response to the reviewers’ comments. We hope that our reply and revision can meet the reviewers’ expectations, and we look forward to the reviewers’ further feedback. Thank you!

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

All raised comments are addressedn in the subsequent sections . Still the author needs to highlight their contribution properly in the abstract as well as in the introduction section .

Now this manuscript good at my end 

Ok