A Lightweight Crop Pest Detection Algorithm Based on Improved Yolov5s

Round 1

Reviewer 1 Report

The paper “A Lightweight Crop Pest Detection Algorithm Based on Im-2 proved Yolov5s” brings an excellent contribution to the agriculture sector by testing a lightweight crop pest detection algorithm based on improved Yolov5s to identify in real-time 8 types of pests on the field. Although the manuscript is written and structured well, I have a few points to ask the authors.

 

1)     How did you determine that the ECMB-Yolov5 model achieved a high detection accuracy while significantly reducing the number of parameters? Did you analyze any trade-offs between accuracy and parameter reduction?

2)     What are the potential implications and benefits of using real-time target detection of crop pests in terms of improving pest control and crop management practices? Have you considered the economic or environmental impact of such a system? I think this is an important point to be considered…

3)     The ongoing works related by the authors “The further work is to analyze the reason for the loss of lightweight model accuracy and optimize the model accuracy, ensure the detection speed and have higher detection accuracy, and finally improve the development of real-time pest detection system.” My question is: Do the authors think about implementing it as an app for smartphones, for example?

 

4)     The agronomic point of view is important to manage the pest on the field. Do you think it's possibly to use UAV images to detect in real-time the pest on the field? Please, provide in the discussion some limitations about the resolution images… 

Author Response

请参阅附件。

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors present a modified YoloV5 architecture for insect detection and classification. Although some of the presented ideas seem to contain some degree of novelty, I think the overall novelty of the paper is not great, since there are numerous publications describing lightweight YoloV5 architectures for insect detection as well as publications on the use YoloV5 + MobileNetV3 backbone + BiFPN applications to other object detection problems.

In my opinion, the brief state-of-the-art presented in section 1 is rather small and incomplete and it fails to clearly position the proposed work on the state-of-the-art. Therefore, I cannot provide a positive paper recommendation at this time.

 

Comments:

1. The state of the art presented in the introduction misses references that closely match the subjects presented in the paper. See for instance (there are more):

Kangshun Li, Jiancong Wang, Hassan Jalil, Hui Wang, “A fast and lightweight detection algorithm for passion fruit pests based on improved YOLOv5,” Computers and Electronics in Agriculture, Vol. 204, 2023, https://doi.org/10.1016/j.compag.2022.107534

Kumar, Nithin, Nagarathna, and Francesco Flammini. 2023. "YOLO-Based Light-Weight Deep Learning Models for Insect Detection System with Field Adaption" Agriculture 13, no. 3: 741. https://doi.org/10.3390/agriculture13030741

Qi, F., Wang, Y., Tang, Z. et al. “Real-time and effective detection of agricultural pest using an improved YOLOv5 network”. J Real-Time Image Proc 20, 33 (2023). https://doi.org/10.1007/s11554-023-01264-0

Zhang Wei, Huang He, Sun Youqiang, Wu Xiaowei, “AgriPest-YOLO: A rapid light-trap agricultural pest detection method based on deep learning”, Frontiers in Plant Science, vol. 13, 2022, https://www.frontiersin.org/articles/10.3389/fpls.2022.1079384

Xiang, Qiuchi, Xiaoning Huang, Zhouxu Huang, Xingming Chen, Jintao Cheng, and Xiaoyu Tang. 2023. "Yolo-Pest: An Insect Pest Object Detection Algorithm via CAC3 Module" Sensors 23, no. 6: 3221. https://doi.org/10.3390/s23063221     

Cheng, Zekai, Rongqing Huang, Rong Qian, Wei Dong, Jingbo Zhu, and Meifang Liu. 2022. "A Lightweight Crop Pest Detection Method Based on Convolutional Neural Networks" Applied Sciences 12, no. 15: 7378. https://doi.org/10.3390/app12157378

My suggestion for the authors is to update the state-of-the-art with closer related (and more recent) work, describing the proposed approaches and later perform comparisons if applicable. The authors should also consider to include a larger “Related work” section since a simple google search reveals far too many closely related work than the ones mentioned in the introduction. YoloV5 improvements to make it “lightweighted” should also be better referenced since the use of MobileNet backbones and Bi-directional Feature Pyramid Network together with Yolo are also ideas already explored in the literature.

2. l61: revise “low detection accuracy, and only detect a single pest, …” – low detection accuracy is vague and may be mischievous if the evaluation datasets used are not the same as the one used in the author’s work – I suggest to summarized the results achieved in the related work as well the datasets used; detection of specific pest species may not necessarily be a negative feature since it may potentially allow the machine learning system to achieve a greater detection performance on species that are indeed relevant for a given crop plant. I suggest rephrasing or further discussing these statements.

3. l65-66: “It is also deployed to the embedded device Jetson Nano [18] for real-time detection.” – from what I understood after reading the paper, the deployment on Jetson Nano was not tested in the field, but in a “lab environment” for computational complexity purposes only. That should be clarified in the introduction.

4. YoloV5 is somewhat outdated – newer Yolo versions exist (e.g., YoloV6, YoloV7, YoloV8, Yolo-NAS). Therefore, in section 2.1 the authors should justify the use of the older YoloV5 as their starting point.

5. The presented approach should be compared with other lightweight YoloV5-based approaches described in the literature under the same conditions.

 

 

The paper reads well.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

In my opinion, the revised manuscript version successfully addressed the major problems that I mentioned in the original submission. The value of the authors’ work is clearer in the revised version and the scope of their work is much better placed in the current state-of-the-art of similar insect detection object detection-based applications.

Regarding the new version, I suggest the following minor tweaks on the manuscript text – the main goals are to improve some explanations and to correct a few glitches.   

 

 General improvements:

In most figures depicting block diagrams I think better image quality could be achieved by using vector formats (pdf, eps, emf, etc.) images. If not, please use raster images with higher DPI.

Some acronyms are not defined - I found some such as CIoU and SIoU, but there may be more – my advice is to perform a careful check in case of paper acceptance.

 

Other corrections/suggestions:

L84 – Define SIoU;

Fig. 1 (as well as figs. 6 and 8) – Why “ConCat” and not “Concat”? Doesn’t it stand for “Concatenation”? Is there a real need to capitalize the 2nd C in "concat"?

L145-L155 – Although the use of Yolov5 is better justified when compared with the first manuscript version, I think that the new version is missing some words showing why the small Yolov5 model (Yolov5s) was chosen instead of the "nano" model (Yolov5n). The nano model would be the one leading to less computational complexity but that is never mentioned.

L182 – Define CIoU;

L220 – “Figure 2 below” -> “Figure 2”;

L225 – “to extract picture features” -> “to extract image features”;

Fig. 3 – “Iutput” -> “Input”; in the block, sometimes capital letter are used, sometimes not – please be consistent (e.g. “Eca_layer” -> “ECA_Layer”);

Fig. 3 – the module “conv_bn_hswish” is presented for the first time here, but the text addressing the figure does not explain what it is;

Table 1 - The meaning of the values on the “Configuration” column are not clear.  My suggestion is to include additional information on the caption or in the text.

Table 1 – Verify the number of parameters on the ECA layer (I was expecting for it to be proportional to the number of input channels on its layer).

L254-L267 – Please improve the explanations – some are a bit confusing – e.g. the text reads “The seven parameters in MobileNet_Block”, but in the table the number of parameters on these blocks varies and does not match seven; the text reads “h_swish activation function”, but the designation doesn’t seem to match the one used in Fig. 3 (“hswish”) .

Fig. 6 – “Iutput” -> “Input”; Check “BiFPN_ConCat2” and “BiFPN_ConCat3” – is there a reason to differentiate? If so, what’s the difference between the “concat2” and “concat3”

L304 – “Angle loss function …” -> “Angle loss function, AL, …” – apply a similar designation for distance loss (DL) and shape loss (SL) on lines 309 and 320.

L348-349 –I think scientific species names usually are designated using italic (e.g., Coccinella Septempunctata) – please check – the same goes for the captions of Fig. 9 and 15 and 16.

L350 – “we randomly divided …” -> “we randomly split …”

L357 – Check eq. (13) – It seems to me that something is wrong/missing – instead of PR chouldn’t it be AP_k (Average Precision for class k)? And in that case a formula or an explanation on computing AP should also be provided.

L415-417 – The explanation regarding [email protected] and [email protected]:0.95, should appear before, in the new paragraph that addresses the results in Table 2.

The English language reads well.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf