Multi-Detection and Segmentation of Potato Seed on Conveyor

Machines with Blur Conditions Using NAFNet and YOLO11

Nurhatinah Hr

and Ingrid Nurtanio

Departement of Informatics Hasanuddin University Makassar, Indonesia

Keywords: Seed Potato, Deep Learning, Nonlinear Activation Free Network (NAFNet), YOLO11.

Abstract: In the agricultural industry, during the potato seed sorting process, automatic seed quality detection is a crucial

requirement for enhancing production efficiency and consistency. This study proposes a potato seed detection

and segmentation system, applicable to both sprouted and unsprouted seeds, by integrating the Nonlinear

Activation Free Network (NAFNet) as the image pre-processing stage and YOLOv11s-Seg as the main

detection model. NAFNet is used to reduce the blurring effect caused by conveyor movement, while

YOLOv11s-Seg is employed to detect and segment potato seed objects. Experiments were conducted using

video data at conveyor speeds of 0.70 m/s, 0.80 m/s, and 0.90 m/s. The evaluation results show that integrating

NAFNet with YOLOv11 significantly improves performance compared to baseline YOLOv11s-Seg and

YOLOv8. At 0.90 m/s, the proposed model achieved an mAP50 of 0.970, precision of 0.941, and recall of

0.890, outperforming YOLOv8, which only reached an mAP50 of 0.880 and a recall of 0.750. Consistent

improvements were also observed at 0.70 m/s, where the system achieved an mAP50 of 0.988, precision of

0.968, and recall of 0.967. NAFNet effectively improves image quality and enhances YOLOv11s-Seg

performance, offering substantial potential for accurate and reliable automation of potato seed sorting.

1 INTRODUCTION

Potatoes are one of the world's strategic horticultural

commodities after rice, wheat, and corn. In Indonesia,

potato productivity is still relatively low at around 59

tons/ha (BPS - Statistics Indonesia, 2024), one of

which is due to the low quality and quantity of seeds.

Sprouting seeds are very important because sprouts

are a key indicator of viability in the process of plant

propagation. However, the bud selection process is

still done manually, which is time-consuming,

inefficient, and error prone.

As technology advances, the application of deep

learning in agriculture is becoming a potential

solution to detect objects automatically and

accurately. Various methods have been developed to

detect potato seed shoots, but they are still limited in

dealing with small objects and sub-ideal

environmental conditions, such as uneven lighting or

ground-like backgrounds (Qiu et al., 2024). This

study aims to improve the detection of water shoots

by using Mask R-CNN and data augmentation. In

https://orcid.org/0009-0006-9737-4974

https://orcid.org/0000-0002-3053-4201

testing, the model with a learning rate of 0.001 and

data augmentation produces an F1-score of 0.966 at a

threshold of 0.8 (Areni et al., 2023) and the research

uses the Mask R-CNN instance segmentation

algorithm with ResNet 101 to detect and classify

coconut shell quality as a raw material for charcoal

briquettes a mean precision value (mAP) of 0.98

(Zikra et al., 2023). Research Gao using the SVM

method and weighted Euclidean distance with a bud

eye recognition accuracy of 91.48%. Although

accurate under controlled conditions, this method is

conventional and cannot work in real-time, making it

less suitable for application to conveyor-based

industrial systems (Gao, 2022). Efforts to improve

performance have been carried out with a deep

learning approach. Li et al., 2025 developed YOLOv8

with a combination of ECA attention, Ghost

convolution, and BiFPN so that it can improve the

detection accuracy of potato shoots with high

efficiency. However, the model focuses more on the

speed of inference and has not yet answered the

Hr, N. and Nurtanio, I.

Multi-Detection and Segmentation of Potato Seed on Conveyor Machines with Blur Conditions Using NAFNet and YOLO11.

DOI: 10.5220/0014266700004928

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Innovations in Information and Engineering Technology (RITECH 2025), pages 17-24

ISBN: 978-989-758-784-9

challenge of detecting very small or blurry objects in

moving images.

Another problem that arises in the industrial

context is the occurrence of motion blur in the image

of potatoes moving on high-speed conveyors. This

phenomenon has a direct impact on the decline in the

accuracy of the detection system. Research gathering

emphasized that blur produced by the movement of

cameras, instruments, and objects is a crucial problem

in the field of precision agriculture, so an image

restoration strategy is needed before the detection

process is carried out (Huihui et al., 2023). Thus,

without an adequate deblurring mechanism, the

detection performance of industrial conveyor systems

will experience significant degradation.

Within this framework, NAFNet is present as one

of the cutting-edge deblurring approaches that is

simple but highly effective as an image restoration

baseline that is able to achieve superior results on

various deblurring and denoising benchmarks with

better computational efficiency than previous

methods (Chen et al., 2022), and NAFNet's

integration with segmentation models, such as

WeedSeg is able to improve the detection of small

objects in low-quality UAV imagery (Genze et al.,

2023).

In the context of the industrial environment, the

main challenge faced is the occurrence of motion blur

due to the movement of conveyors at high speeds,

which has an impact on the decrease in the accuracy

of the detection system. Therefore, an approach is

needed that not only has a high level of accuracy but

is also capable of handling blurry images. One

strategy that can be applied is to integrate image

restoration models such as NAFNet, which functions

to improve image quality through the deblurring

process, before the detection stage is carried out.

Furthermore, the object identification process can be

carried out using modern detection models such as

YOLOv11, which are able to produce more precise

detection and segmentation. The combination of these

two methods is expected to improve the system's

resistance to visual interference as well as improve

the accuracy of potato bud detection in high-speed

conveyor conditions.

2 MATERIALS AND METHODS

This research focuses on the development of multiple

detection and segmentation of potato seeds in

conveyor machines with fuzzy conditions using deep

learning algorithms. The research workflow is

illustrated in Figure 1.

Figure 1: Research Process Workflow.

Figure 1 illustrates the proposed system design for

segmenting potato seeds in blurry conditions at

conveyor speeds of 0.70 m/s, 0.80 m/s, and 0.90 m/s,

which can be divided into several main stages as

follows:

2.1 Data Acquisition

The data collection process in this study was obtained

from real images of sprouted and unsprouted potato

seeds. The data was collected from video footage of

potato seeds moving on a conveyor with a Logitech

camera that has a resolution of 1080p and a frame rate

of 30 or 60 frames per second with a camera height of

15 cm. The data collection process was carried out at

three conveyor speeds, namely 0.70 m/s, 0.80 m/s,

and 0.90 m/s, to reflect the expected operational

variations in the industry.

Figure 2: Data collection techniques.

Figure 3: Classification of potato seeds:

(a) sprouted and

(b) non-sprouted.

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

(a) (b) (c)

Figure 4: Input Image with Conveyor Speeds. (a) 0,70 m/s,

(b) 0,80 m/s, (c) 0,90 m/s.

The study also included variations of potato seed

images taken at different levels of conveyor speeds in

the dataset. This approach aims to reflect more

realistic operational conditions as encountered in the

sorting process in an industrial environment. Thus,

the developed model is expected to have a higher

level of accuracy and reliability in detecting and

classifying potato seeds. In addition, this diversity of

scenarios allows the model testing and validation

process to be carried out more comprehensively, so

that model performance can be objectively assessed

against various real-world environmental conditions.

2.2 Pre-Processing

The dataset obtained has gone through an annotation

process using polygon labeling, with class categories

consisting of two main labels, namely sprouted and

non-sprouted, as shown in Figure 3. Considering that

the size of the acquired image varies, an image

resizing process is carried out so that all images have

uniform dimensions of 640 × 640 pixels.

Furthermore, the labeled datasets are divided into

three subsets, namely the train set, validation set, and

test set, with a distribution ratio of 70:20:10

sequentially. From this process, 1,365 images were

obtained for training, 390 images for validation, and

195 images for testing. The data distribution for each

class is as follows:

Table 1: Class Distribution in Potato Seeds.

Dataset S

route

Non-S

route

Train 3093 3636

Vali

1460 1515

Test 653 713

This distribution shows that the amount of data

between classes is relatively balanced in each subset,

with a small distribution difference (<10%). This

indicates that the dataset does not experience

significant class imbalance problems, so it can

support the model training process in a more

representative manner and reduce the risk of bias

towards one of the classes. Thus, the dataset used is

of sufficient quality to support the segmentation and

classification experiment of sprouted and non-

sprouted potato seeds. In addition to the annotation

and image size alignment process, the pre-processing

stage also includes improving image quality using the

Nonlinear Activation Free Network (NAFNet) as an

image restoration step. The application of NAFNet

aims to reduce the effect of motion blur arising from

the movement of potato seeds on the conveyor,

resulting in a sharper and more representative image.

Thus, the quality of the dataset is significantly

improved, while strengthening the accuracy of the

detection and segmentation model at the next stage of

development.

2.3 Model Development & Training

2.3.1 Nonlinear Activation Free Network

(NAFNet)

NAFNet represents an innovative deep learning

architecture. NAFNet has demonstrated outstanding

achievements in image recovery applications,

including addressing challenges in blurry image

recovery, noise reduction, and stereo super-

resolution. Specifically designed to address problems

at the pixel level, NAFNet innovatively replaces

batch normalization with layer normalization at the

pixel level to improve pixel accuracy. The use of U-

Net structures and jump connections in NAFNet

facilitates better transfer of information from input to

output, preventing the decline in accuracy associated

with information loss (Maruzuki et al., 2024). In the

context of our research on potato seeds moving over

a conveyor in classifying between sprouting and non-

sprouting, NAFNet's dense prediction capabilities at

the pixel level hold promise for accurately describing

features in regions of interest (ROI), demonstrating

potential applications in improving the accuracy of

object detection, especially in challenging image-

making conditions such as blurred images of potato

seeds. The NAFNet subblocks are shown in Figure 5.

Figure 5: NAFNet Subblock.

The formulation of Gated Linear Units (GLU) is as

follows:

𝐺𝑎𝑡𝑒(𝑋, 𝑓, 𝑔, 𝜎) = 𝑓

(

𝑥

)

⨀𝜎(𝑔

(

𝑋

)

) (1)

Multi-Detection and Segmentation of Potato Seed on Conveyor Machines with Blur Conditions Using NAFNet and YOLO11

where X shows a feature map, f and is a linear

transformer by representing a non-linear activation

function such as Sigmoid, along with 𝑔𝜎⊙ showing

element-by-element multiplication, has shown

potential to improve performance. However, it should

be noted that the introduction of GLU has led to an

unexpected increase in intra-block complexity.

The Gaussian Error Linear Unit (GELU) is

expressed by the following formula:

𝐺𝐸𝐿𝑈

(

𝑥

)

=𝑥∅ (𝑥) (2)

where it shows the cumulative distribution

function of the standard normal distribution, it should

be emphasized that GELU inherently integrates

nonlinearity and is free of activation functions.

Therefore, based on this concept, the feature map is

divided into two segments along the channel

dimensions, facilitating the multiplication of channels

and elements. A simplified schematic representation

of this gate is shown below, and its mathematical

expressions are provided below: ∅𝜎

𝑆𝐺

(

𝑋, 𝑌

)

= 𝑋 ⨀ 𝑌 (3)

where X and Y are feature maps of the same size.

Additionally, this paper integrates simplified channel

attention mechanisms into nonlinear networks

without activation. Spatial information undergoes

initial compression into the channel. The duct

attention mechanism has similarities to the linear unit

of the Berber and can be considered a special form of

the GLU. By maintaining the global essence of

information aggregated in channel attention and

facilitating interaction with channel-specific

information, formally simplified channel attention is

articulated as:

𝑆𝐶𝐴

(

𝑋

)

= 𝑋 ∗ 𝑊 𝑝𝑜𝑜𝑙 (𝑋) (4)

where X represents a feature map, the pool denotes

the grouping of the global average applied to the

spatial information consolidated into the channel, and

∗ denotes the channel multiplication operation.

2.3.2 YOLO (You Only Look Once)

The study leveraged the YOLOv8 and YOLOv11

models for object detection, both of which are part of

the YOLO (You Only Look Once) family, which is

renowned for its real-time detection with a high level

of accuracy.

YOLOv8: YOLOv8 presents a modular

structure consisting of Backbone, Neck, and Head

structures, supporting multi-scale feature

representation for the recognition of small objects.

The Neck Module combines the Feature Pyramid

Network (FPN) and the Path Aggregation Network

(PANet) to improve object recognition at various

scales by integrating the characteristics of multiple

layers. The architecture also includes spatial pyramid

pooling (SPPF) for more efficient spatial data storage

(Jocher et al., 2023)

Figure 6: The Architecture diagram of YOLOv8.

2) YOLOv11: YOLO11, the most recent iteration

of the YOLO series, is a high-performance, low-

latency object detection method founded on

convolutional neural networks (CNN). YOLO11, in

contrast to its predecessors, incorporates an enhanced

backbone and neck architecture, which augments its

feature extraction proficiency, facilitating elevated

mAP and expedited inference rates. The YOLO11

series includes five different architectures:

YOLO11n, YOLO11s, YOLO11m, YOLO11l, and

YOLO11x. The primary distinctions among these

variants lie in the configuration of feature extraction

modules and convolutional filters at specific locations

within the network. Both model size and parameter

count increase progressively from YOLO11n to

YOLO11x (Zhao & Jiang, 2025). The introduction of

Cross Stage Partial with Spatial Attention (C2PSA)

blocks and other architectural improvements has

improved YOLOv11's real-time detection capabilities

while maintaining computational efficiency.

YOLOv11 also optimizes the Neck structure,

allowing for more efficient feature extraction and

faster convergence (Jocher et al., 2024).

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

Figure 7: The Architecture of YOLO11.

Figure 8: YOLO11 builds on YOLOv8 architecture with

refined detection and segmentation efficiency and accuracy

over the benchmark dataset.

Table 2 shows the hyperparameter adjustment

configuration used in the model training process.

Table 2: Hyperparameter Adjustments.

Hyperparamete

Value

Image Size 640 x 640

Epochs 300

Momentum 0.937 (SDG)

weight_decay 0.0005

Learning Rate 0.01

2.4 Model Evaluation

The model evaluation approach was carried out to

assess the effectiveness in detecting and categorizing

potato seed entities. The performance measurement

method utilizes commonly used evaluation

methodologies, specifically precision, recall, and

mAP. The Confusion Matrix is obtained based on

calculations using equations (5), (6), (7) as follows:

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =





(5)

Precision measurement measures the proportion

of correctly detected potato seeds from all the

predictions the model produces.

𝑅𝑒𝑐𝑎𝑙𝑙 =





(6)

Recall measures the proportion of potato seeds

that are successfully detected correctly compared to

all potato seeds that are actually present at ground

truth.

𝑚𝐴𝑃 =





∑

𝐴𝑃







(7)

Where represents the Average Precision value for

the second class.𝐴𝑃



i, while N indicates the total

number of classes. Mean Average Precision (mAP) is

widely used in the field of object detection as a key

metric to assess the overall performance of the model

across the entire evaluated class (Li et al., 2025b).

For the complexity aspect of the model, there are

three metrics: Params, GFlops, and Size, which are

calculated according to equations (8) and (9):

𝑃𝑎𝑟𝑎𝑚𝑠 =



𝑟 x

(

𝑓 x 𝑓

)

x 𝑜



𝑜 (8)

𝐹𝑙𝑜𝑝𝑠 = 𝑜

(∑

𝑘







x 𝐶





x 𝐶





∑

𝑀



x 𝐶







)

(9)

3 RESULTS AND DISCUSSION

3.1 Model Training Configuration

In the results of the experiment, the impact of

hyperparameter combinations on the model was

evaluated. The main metrics used for comparison are

Mean Average Precision (mAP), Precision and

Recall. The following sections present a quantitative

and qualitative analysis of these metrics, comparing

the performance of YOLOv8-seg and YOLOv11-seg.

Multi-Detection and Segmentation of Potato Seed on Conveyor Machines with Blur Conditions Using NAFNet and YOLO11

Table 3: Comparison of YOLOv8s-seg and YOLOv11s-seg

Models on Various Metrics.

Metric

Model

YOLOv8 YOLOv11

YOLOv11

with

NAFNet

mAP 50 0,938 0,942 0,948

mAP 50-95 0,91 0,919 0,928

Accurac

0,861 0,881 0,894

Recall 0,898 0,901 0,902

Speed (GPU) 3.4 ms 3.4 ms 3.4 ms

Params 11 M 10 M 10 M

FLOPs 39.9 B 35.3 B 35 B

Table 3 shows that YOLOv11s have performance

advantages over YOLOv8s, especially in terms of

detection accuracy, precision, and recall, while

maintaining computational efficiency. The

application of NAFNet on YOLOv11s-Seg further

improves model performance, especially in

overcoming images that experience degradation due

to blur from the movement of objects on the

conveyor.

3.2 Segmentation Potato Seed Testing

The model was tested using a test dataset to verify its

performance in detecting, segmenting, and

classifying potato seeds under real conditions. In the

early stages, NAFNet was applied as an image

restoration method to reduce the effect of motion blur

that arises due to the movement of seeds on the

conveyor. The test was carried out under various

conditions of conveyor speed, namely 0.70 m/s, 0.80

m/s, and 0.90 m/s, which significantly caused

variations in the blur level in the test image. After

going through the restoration process, the image is

then processed by YOLOv11-seg to detect objects,

segment them, and classify potato seeds into two

main categories, namely Sprouted and Non-Sprouted.

Table 4: Model Test Results.

Method

Speed

(m/s)

mAP 50

(%)

P (%) R (%)

YOLOv8s-Seg

0,70 0,91 0,93 0,790

0,80 0,90 0,838 0,770

0,90 0,88 0,830 0,750

YOLOv11s-Seg

0,70 0,925 0,951 0,794

0,80 0,908 0,86 0,789

0,90 0,873 0,842 0,754

YOLOv11s-Seg

with NAFNet

0,70 0,98 0,968 0,967

0,80 0,972 0,955 0,90

0,90 0,97 0,941 0,89

Based on the test results in Table 4, it can be seen

that the YOLOv11s-Seg model consistently shows

better performance than YOLOv8s-Seg across all

conveyor speed variations. For example, at a speed of

0.90 m/s, YOLOv11s-Seg obtained an mAP value of

0.873 with a precision of 0.842, surpassing

YOLOv8s-Seg, which only achieved an mAP of 0.88

with a precision of 0.830. This finding confirms that

the YOLOv11s-Seg architecture has an advantage in

maintaining detection accuracy at high speeds.

Furthermore, the integration of NAFNet as a pre-

processing stage resulted in a significant performance

improvement. At a speed of 0.90 m/s, the

combination of YOLOv11s-Seg with NAFNet

achieved an mAP of 0.970 with a precision of 0.941

and a recall of 0.890, which is substantially higher

than the two previous models. This upward trend is

consistent across all speeds; for example, at a speed

of 0.70 m/s, the mAP increases to 0.988 with a

precision of 0.968 and a recall of 0.967. This shows

that the application of NAFNet can effectively

overcome image quality degradation due to blurring

effects on the conveyor, thereby maintaining model

performance stability.

Overall, the experimental results show that the

integration of NAFNet with YOLOv11s-Seg is

capable of producing an optimal combination of

accuracy, precision, and recall. This makes the

approach more reliable in supporting the detection

and segmentation of potato seeds under various

conveyor speed conditions. Visualization of the

detection results in the test image shows that

YOLOv11s-Seg, after undergoing pre-processing

with NAFNet, can accurately recognize and segment

potato seed objects. This is indicated by the

appearance of bounding boxes and segmentation

areas that correspond to the position of the objects.

The role of NAFNet is proven to be significant in

reducing the blur effect that occurs due to conveyor

movement, so that the contours and structure of the

seeds appear clearer. Thus, the application of

NAFNet not only improves the quality of the input

image but also directly contributes to improving the

performance of YOLOv11s-Seg in detecting and

segmenting potato seeds at varying conveyor speeds.

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

(a) (b)

Figure 9: Testing on real images, conveyor speed 0.70 m/s

with Model (a) YOLOv11 Baseline and (b) YOLOv11 with

NAFNet.

(a) (b)

Figure 10: Testing on real images, conveyor speed 0.80 m/s

with Model (a) YOLOv11 Baseline and (b) YOLOv11 with

NAFNet.

(a) (b)

Figure 11: Testing on real images, conveyor speed 0.90 m/s

with Model (a) YOLOv11 Baseline and (b) YOLOv11 with

NAFNet.

4 CONCLUSIONS

This study proposes an automatic detection and

segmentation system based on YOLOv11s-Seg

combined with NAFNet as a pre-processing stage to

overcome image quality degradation caused by

conveyor movement. Experimental evaluation results

show that the integration of NAFNet consistently

improves model performance across all conveyor

speed variations. This improvement is reflected in the

increase in mAP, precision, and recall values

compared to YOLOv8s-Seg and YOLOv11s-Seg

without pre-processing.

At a conveyor speed of 0.90 m/s, the combination

of YOLOv11s-Seg with NAFNet achieved an mAP

of 0.970, precision of 0.941, and recall of 0.890,

surpassing the performance of YOLOv8s-Seg, which

only achieved an mAP of 0.88 with a precision of

0.830 under the same conditions. A similar upward

trend was also observed at a speed of 0.70 m/s, where

the integration of NAFNet resulted in an mAP of

0.988, precision of 0.968, and recall of 0.967, which

was significantly higher than the model without pre-

processing. Thus, the application of NAFNet not only

serves to improve the quality of the input image but

also proves to strengthen the robustness and stability

of YOLOv11s-Seg detection in a dynamic industrial

environment with varying conveyor speeds.

Although the results obtained show a significant

improvement in performance, this study still has a

number of limitations. The evaluation was conducted

on a dataset with limited coverage and relatively

controlled environmental conditions, so the system's

ability to generalize to real-world situations with

higher complexity still needs to be further validated.

Furthermore, although the integration of NAFNet has

been proven to improve image quality and does not

cause a significant decrease in inference time, further

optimization is still needed so that the system can be

applied efficiently on a large scale and in real-time

applications with high computational loads. For

further research, testing in actual industrial

environments, development of model architecture

optimization strategies—for example, based on

attention mechanisms—and expansion of the dataset

are needed so that the system can be more adaptive to

varying field conditions. This direction of

development is expected to improve the efficiency

and accuracy of detection, thereby strengthening the

potential for applying computer vision technology in

the automatic selection and segmentation of potato

seeds in the modern agricultural sector.

ACKNOWLEDGEMENTS

The authors would like to thank the Master’s Program

in Informatics Engineering at Hasanuddin University

and the Artificial Intelligence and Multimedia

Processing (AIMP) Thematic Research Group for

their support and facilities during this research. The

authors also acknowledge colleagues who provided

constructive feedback and assistance throughout the

research process. Additionally, the authors

acknowledge the use of a generative AI tool to

enhance the clarity and grammar of this manuscript.

All contents, analyses, and conclusions remain the

sole responsibility of the authors.

Multi-Detection and Segmentation of Potato Seed on Conveyor Machines with Blur Conditions Using NAFNet and YOLO11

REFERENCES

Areni, I. S., Maulidyah, N., Indrabayu, Bustamin, A., &

Arief, A. B. (2023). Increasing Precision of Water

Sprout Detection based on Mask R-CNN with Data

Augmentation. International Journal on Advanced

Science, Engineering and Information Technology,

13(2), 794–800.

https://doi.org/10.18517/ijaseit.13.2.16468

BPS - Statistics Indonesia. (2024). Produksi Tanaman

Sayuran - Tabel Statistik - Badan Pusat Statistik

Indonesia. https://www.bps.go.id/id/statistics-

table/2/NjEjMg==/produksi-tanaman-sayuran.html

Chen, L., Chu, X., Zhang, X., & Sun, J. (2022). Simple

Baselines for Image Restoration. Lecture Notes in

Computer Science (Including Subseries Lecture Notes

in Artificial Intelligence and Lecture Notes in

Bioinformatics), 13667 LNCS, 17–33.

https://doi.org/10.1007/978-3-031-20071-7_2

Gao, S. (2022). Research on detection method of sprouted

potato based on SVM and weighted Euclidean distance.

6th International Conference on Mechatronics and

Intelligent Robotics, 92.

https://doi.org/10.1117/12.2644666

Genze, N., Wirth, M., Schreiner, C., Ajekwe, R., Grieb, M.,

& Grimm, D. G. (2023). Improved weed segmentation

in UAV imagery of sorghum fields with a combined

deblurring segmentation model. Plant Methods, 19(1).

https://doi.org/10.1186/S13007-023-01060-8

Huihui, Y., Daoliang, L., & Yingyi, C. (2023). A state-of-

the-art review of image motion deblurring techniques in

precision agriculture. Heliyon, 9(6).

https://doi.org/10.1016/j.heliyon.2023.e17332

Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics

yolov8.

Jocher, G., Qiu, J., & Chaurasia, A. (2024). Ultralytics

yolo11. GitHub Repository.

Li, Y., Zhao, Q., Zhang, Z., Liu, J., & Fang, J. (2025a).

Detection of Seed Potato Sprouts Based on Improved

YOLOv8 Algorithm. Agriculture 2025, Vol. 15, Page

1015, 15(9), 1015.

https://doi.org/10.3390/AGRICULTURE15091015

Li, Y., Zhao, Q., Zhang, Z., Liu, J., & Fang, J. (2025b).

Detection of Seed Potato Sprouts Based on Improved

YOLOv8 Algorithm. Agriculture (Switzerland), 15(9).

https://doi.org/10.3390/agriculture15091015

Maruzuki, M. I. F., Osman, M. K., Shafie, A. S., Setumin,

S., Ibrahim, A., Saleh, H. M., Tahir, M. S. M., &

Rabiain, A. H. (2024). Road Image Deblurring with

Nonlinear Activation Free Network. 14th IEEE

International Conference on Control System,

Computing and Engineering, ICCSCE 2024 -

Proceedings, 288–293.

https://doi.org/10.1109/ICCSCE61582.2024.10696495

Qiu, Z., Wang, W., Jin, X., Wang, F., He, Z., Ji, J., & Jin,

S. (2024). DCS-YOLOv5s: A Lightweight Algorithm

for Multi-Target Recognition of Potato Seed Potatoes

Based on YOLOv5s. In Agronomy (Vol. 14, Issue 11).

https://doi.org/10.3390/agronomy14112558

Zhao, Y., & Jiang, Z. (2025). YOLO-WWBi: An Optimized

YOLO11 Algorithm for PCB Defect Detection. IEEE

Access

, 13(May), 74288–74297.

https://doi.org/10.1109/ACCESS.2025.3564734

Zikra, A. A., Ilham, A. A., Nurtanio, I., & Saptadi, N. T. S.

(2023). Multi Detection and Segmentation Coconut

Shell for Charcoal Briquette Using Mask R-CNN. 2023

International Seminar on Intelligent Technology and Its

Applications: Leveraging Intelligent Systems to

Achieve Sustainable Development Goals, ISITIA 2023 -

Proceeding, c, 615–620.

https://doi.org/10.1109/ISITIA59021.2023.10221025

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology