Available via license: CC BY 4.0
Content may be subject to copyright.
RESEARCH ARTICLE
A method for detecting the quality of cotton
seeds based on an improved ResNet50 model
Xinwu DuID
1,2☯
*, Laiqiang Si
1☯
, Pengfei Li
1☯
, Zhihao Yun
1☯
1College of Agricultural Equipment Engineering, Henan University of Science and Technology, Luoyang,
Henan, China, 2Science & Technology Innovation Center for Completed Set Equipment, Longmen
Laboratory, Luoyang, Henan, China
☯These authors contributed equally to this work.
*du_xinwu@sina.com.cn
Abstract
The accurate and rapid detection of cotton seed quality is crucial for safeguarding cotton cul-
tivation. To increase the accuracy and efficiency of cotton seed detection, a deep learning
model, which was called the improved ResNet50 (Impro-ResNet50), was used to detect cot-
ton seed quality. First, the convolutional block attention module (CBAM) was embedded into
the ResNet50 model to allow the model to learn both the vital channel information and spa-
tial location information of the image, thereby enhancing the model’s feature extraction
capability and robustness. The model’s fully connected layer was then modified to accom-
modate the cotton seed quality detection task. An improved LRelu-Softplus activation func-
tion was implemented to facilitate the rapid and straightforward quantification of the model
training procedure. Transfer learning and the Adam optimization algorithm were used to
train the model to reduce the number of parameters and accelerate the model’s conver-
gence. Finally, 4419 images of cotton seeds were collected for training models under con-
trolled conditions. Experimental results demonstrated that the Impro-ResNet50 model could
achieve an average detection accuracy of 97.23% and process a single image in 0.11s.
Compared with Squeeze-and-Excitation Networks (SE) and Coordination Attention (CA),
the model’s feature extraction capability was superior. At the same time, compared with
classical models such as AlexNet, VGG16, GoogLeNet, EfficientNet, and ResNet18, this
model had superior detection accuracy and complexity balances. The results indicate that
the Impro-ResNet50 model has a high detection accuracy and a short recognition time,
which meet the requirements for accurate and rapid detection of cotton seed quality.
1. Introduction
Cotton seed is the foundation of cotton production, and its quality directly impacts cotton
yield and quality [1,2]. The quality of cotton seed is under increasing scrutiny as the mecha-
nized one-hole, one-seed precision sowing technology becomes more prevalent in China [3–
5]. Phenotypic defects are one of the criteria for evaluating the quality of cotton seed. Cotton
seed defects are traditionally detected manually, which is laborious, time-consuming, and sub-
jective. Therefore, developing an objective and automated method for detecting cotton seeds is
necessary.
PLOS ONE
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 1 / 19
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
OPEN ACCESS
Citation: Du X, Si L, Li P, Yun Z (2023) A method
for detecting the quality of cotton seeds based on
an improved ResNet50 model. PLoS ONE 18(2):
e0273057. https://doi.org/10.1371/journal.
pone.0273057
Editor: Sathishkumar V. E., Hanyang University,
REPUBLIC OF KOREA
Received: January 25, 2022
Accepted: July 28, 2022
Published: February 15, 2023
Peer Review History: PLOS recognizes the
benefits of transparency in the peer review
process; therefore, we enable the publication of
all of the content of peer review and author
responses alongside final, published articles. The
editorial history of this article is available here:
https://doi.org/10.1371/journal.pone.0273057
Copyright: ©2023 Du et al. This is an open access
article distributed under the terms of the Creative
Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in
any medium, provided the original author and
source are credited.
Data Availability Statement: All relevant data are
within the manuscript and its Supporting
Information files.
Funding: Funding:National Natural Science
Foundation of China (52075150), Natural Science
Machine learning-based image processing techniques have been successfully applied to
detect seed quality with the advancement of computer vision technology [6–8]. The research-
ers conduct seed quality assessment by extracting features such as texture, color and shape of
the seed images. This method is more advanced and effective in detecting seed quality than the
manual method. However, the method is relatively dependent on manual feature extraction,
and different features require different extraction methods. In addition, manual feature extrac-
tion is usually inadequate. Thus, it leads to the detection accuracy of the method is not high.
There has been an increase in convolutional neural networks (CNN) used for image recog-
nition [9–11]. In addition to simulating the human brain’s mechanism for extracting features
in layers, the technique can extract features automatically from simple to complex, from bot-
tom to top, and from concrete to abstract. Several researchers have successfully applied CNN
to the detection of seed quality [12–15]. But, a disadvantage of CNN detection is that it
requires a large amount of training data, is time-consuming, and is computationally resource-
intensive.
To address the shortcomings of existing methods, this paper proposes a new CNN for cot-
ton seed quality detection. A summary of this study’s major contributions and innovations is
provided below.
1. Based on the appearance of defects in cotton seed, a new cotton seed dataset is created to
support the development of subsequent detection algorithms.
2. The Impro-ResNet50 model is proposed as a new method for detecting cotton seed quality
based on an attention mechanism. The CBAM attention block is embedded in ResNet50 to
integrate feature channel and spatial information attention and enhance the model’s capac-
ity to learn essential information about cotton seed regions.
3. The model’s application serves as a reference for developing new models, demonstrating
the interoperability of deep learning models and attention mechanisms.
4. On the basis of the cotton seed quality identification dataset, Impro-ResNet50 is subjected
to extensive comparative experiments. Impro-ResNet50 is highly accurate and robust in
cotton seed detection tasks, demonstrating the efficacy of the CBAM module. Provide tech-
nical support for developing cotton seed quality testing equipment in the future.
2. Related works
2.1. Application of machine vision technology to the detection of seed
quality
The machine vision-based detection technology of seed quality has become relatively mature.
Using image processing technology, the authors of [16] created an online detection system for
soybean seeds. The system was based on classifying surface information such as the color, tex-
ture, and shape of soybeans and achieved a detection accuracy of over 97% for cracked and
healthy soybeans. The authors of [17] chose high-quality pepper seeds using machine vision
and classifiers. Multiple physical characteristics, such as the seeds’ width, length, and projected
area, were used as classification criteria. It detected high-quality and low-quality seeds with a
greater than 90% accuracy. The authors of [18] described a low-rank Joint Multi-Modal bag-
of-feature (JMBoF) classification method for detecting the appearance quality of soybean
seeds. The model achieved 82.1% accuracy in detecting healthy, good, and unhealthy soybean
seeds based on the color of the seeds. The authors of [19] combined spectral imaging and
machine vision techniques to detect damage to sugar beet seeds. This method achieved a detec-
tion accuracy of 82% for five distinct types of sugar beet damage. In [20], the authors proposed
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 2 / 19
Foundation of Henan Province (No.202300410124)
and Guangdong Key R&D Program
(No.2019B020222004).
Competing interests: The authors have declared
that no competing interests exist.
a machine vision-based, one-class classification method for evaluating the quality of tomato seeds.
A 97% accuracy rate was achieved in classifying healthy and infected seeds. Combining automatic
X-ray analysis and machine learning models, the authors of [21] presented a method for classify-
ing the quality of Jatropha curcas seeds. The technique detected normal and abnormal seeds with
a 94.36% accuracy rate. In [22], the authors developed a machine vision-based algorithm to detect
moldy and normal maize seeds based on the difference in surface color, which had an overall
detection accuracy of no less than 94%. In [23], the authors developed a machine vision-based
double-sided rice seed identification system. The method identified rice seeds with open glumes
using Hough linear detection and feature extraction. The algorithm achieved recognition accura-
cies of 88.1% and 87.7%, respectively, for normal and open rice seeds.
Although the above methods achieve excellent seed quality detection performance, it requires
cumbersome image pre-processing and feature extraction. In addition, the input feature data lim-
ited the model’s accuracy, which was often inadequate, resulting in poor detection accuracy.
2.2. Application of convolutional neural networks to the detection of seed
quality
The CNN has started to be used to perform the quality detection work of seeds. For instance,
the authors of [24] demonstrated a CNN-based transfer learning method for detecting haploid
and diploid maize seeds. The model achieved optimal detection accuracy of 94.22%, providing
technical support for the non-destructive, rapid, and inexpensive detection of high-quality
seeds. In [25], the authors developed a peanut seed quality detection method based on machine
vision and an adaptive CNN. The process achieved an average detection accuracy of 99.70% for
common peanut seeds, such as mouldy, broken, or shrivelled. The authors of [26] integrated
near-infrared hyperspectral imaging (NIR-HSI) and CNN deep learning techniques to differen-
tiate between viable and inviable seeds. The process achieved a 90% detection rate for seeds. In
[27], the authors presented an enhanced MobileNetV2-based model for detecting soybean seeds
of superior quality. The detection accuracy of this model was 97.84%, which achieved the best
results compared to the other seven models mentioned in the paper for detecting the quality of
soybean. The authors of [28] claim that a photonic sensor based on laser backscattering and
deep transfer learning was used to detect seeds of superior quality. The method achieved a
98.31% detection rate for high and low-quality seeds. Based on deep convolutional generative
adversarial networks (DCGAN) and NIR-HSI, the authors of [29] proposed a method for iden-
tifying substandard wheat. In comparing support vector machine (SVM) and decision tree
(DT) classifiers, the method demonstrated the best performance, with 96.67% detection accu-
racy for unsound wheat. Another study [30] presented a model for detecting maize seed defects
based on a watershed algorithm and a dual-pathway CNN model. This method outperformed
the conventional image processing techniques mentioned in the paper, with an average detec-
tion accuracy of 95.63% for both defective and healthy maize seeds.
Although seed quality detection has been extensively studied in previous research, there are
currently no mature CNN-based detection models for cotton seed quality detection. Conse-
quently, we anticipate that the proposal will address the current limitations of cotton seed
quality detection models and reduce costs without compromising detection performance.
3. Experimental data
3.1. Data acquisition
GK-10 lazy cotton seeds harvested in 2021 were utilized as experimental material. This cotton
seed variety was widely cultivated, high-yielding, disease-resistant, well-adapted, and
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 3 / 19
representative. A random sample of 50 copies from the purchased material was taken, and 100
cotton seeds were randomly selected from each copy for integrity detection. The results
showed that the proportion of cracked and partially broken cotton seeds in the material ranged
from 5% to 10%. The phenotypic characteristics of the three types of cotton seeds were deter-
mined by observing intact, broken, and cracked cotton seeds in the material. Overall, intact
cotton seeds were brown with entire edges and no discernible surface defects. Cracked cotton
seeds were distinguished by surface cracks and a shift in color gradation at the cracks. Broken
cotton seeds exposed the milky white endosperm at their edges.
In the indoor environment (Natural Light + energy-efficient lamp), each batch of 18 seeds
was distributed randomly in a 36 pattern. The seed samples were photographed vertically from
20 to 25 cm using a Hikvision CCD (MV-CE200-10UC model) camera and 12 mm lens
(MVL-HF1224M-10MP model) with an image resolution of 4024×3072 pixels. 3154 images
were acquired in total. The image acquisition system is shown in Fig 1.
The image of single cotton seed was produced by cropping the entire picture. To meet the
image input requirements of the CNN, the cotton seed image was uniformly scaled to 224×224
pixels. The individual seed images and the corresponding decomposition background images
are shown in Fig 2. A total of 4419 images of cotton seeds were obtained, consisting of 1367
intact seeds, 1467 cracked seeds, and 1585 broken seeds.
3.2. Data preprocessing
To improve the model’s generalisation and robustness, the data were expanded by flipping,
rotating, scaling, cropping, panning, and adding noise to the three image types of cotton seed.
The expanded cotton seed image dataset consisted of 7386 images, and the dataset was divided
into 80% training set, and 20% validation set randomly using the Python program. The sample
distribution of cotton seeds is shown in Table 1 (S1 Data).
4. Methodology
In cotton cultivation, low-quality cotton seeds lead to a reduction in yield and quality. At
this time, deep learning techniques can detect cotton seed quality early and avoid sowing
low-quality cotton seeds. To effectively detect the quality of cotton seeds, a deep learning
network based on residual structure and embedded attention mechanism was proposed in
this paper.
Fig 1. Cotton seed acquisition. (a) Image acquisition system: 1. Camera, 2. Lens, 3. Light-emitting diode (LED) lamps, 4. Cotton seed, 5. Platform, 6.
Image monitor. (b) Cotton seed image.
https://doi.org/10.1371/journal.pone.0273057.g001
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 4 / 19
4.1. Model improvement and structural design
4.1.1. ResNet50 network structure. When the CNN depth is increased, gradient degrada-
tion and disappearance will occur during the training process, resulting in difficulty in conver-
gence and low accuracy [31,32]. However, adding a residual structure to the CNN can largely
avoid this phenomenon. The ResNet50 model and residual structure are shown in Fig 3.
The structure of the ResNet50 model is shown in Fig 3A. In Stage 1, the input image was
reduced in size using a 7×7 convolutional layer and 3×3 maximum pooling downsampling.
Then the higher-level features were extracted using the Conv2, Conv3, Conv4, and Conv5
residual structures in Stage 2. As a final step, the extracted high-dimensional features were fed
into the fully-connected layer of Stage 3 for classification.
Two types of structures are available for the residual block, as shown in Fig 3B and 3c.
Using 1×1 convolutional kernels before and after the 3×3 convolutional kernels to downscale
and upscale could reduce the number of parameters in the model. Residual structure 3b was
Fig 2. Single cotton seed.
https://doi.org/10.1371/journal.pone.0273057.g002
Table 1. Cotton seed data.
Category Data set Training set Validation set
Intact seed 2367 1894 473
Broken seed 2465 1972 493
Cracked seed 2554 2044 510
https://doi.org/10.1371/journal.pone.0273057.t001
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 5 / 19
the block with added scale, where the output feature matrix’s height and width were half of the
input through shortcut branching. This operation contributed to preventing model degrada-
tion. The residual structure 3c indicated that the feature size was unaltered, indicating that the
output feature matrix’s height and width were also unaltered.
4.1.2. Convolutional block attention module model. The CBAM module is a highly effi-
cient attention module that can be incorporated quickly and flexibly into conventional classifi-
cation networks without adding a large number of parameters, thereby enhancing the
representation of features in convolutional neural networks [33,34]. Using the CBAM module,
The ResNet50 model could extract the features of cotton seed image channels while retaining
the property of accurate spatial location information. The structure of CBAM is shown in
Fig 4.
The Channel Attention Mechanism and Spatial Attention Mechanism made up the CBAM
module. Given an input feature F, a channel compression operation was used to generate the
channel attention weight M
C
. Then, M
C
was multiplied by Fto obtain F0. The spatial attention
weight M
S
was then generated by a 2-dimensional spatial compression operation and multi-
plied by F0to produce F00. The specific calculation process is given in Eq 1.
F0¼MCðFÞ F
F@¼MSðF0Þ F0ð1Þ
(
Fig 3. The network structure of the Resnet50 model.
https://doi.org/10.1371/journal.pone.0273057.g003
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 6 / 19
where F2RCHWrepresents the input feature matrix. F02RCHWrepresents the feature
mapping selected by channel attention. F@represents the feature mapping selected by spatial
attention. represents the element multiplication. MC2RC11and MS2R1HWrepresents
the channel attention weights, and the spatial attention weights, respectively. The calculations
of M
C
and M
S
are given in Eqs 2and 3.
MCðFÞ ¼ sðMLPðAvgPoolðFÞÞ þ MLPðMaxPoolðFÞÞÞ ð2Þ
MSðFÞ ¼ sðf77ð½AvgPoolðFÞ;MaxPoolðFÞÞÞ ð3Þ
where MLP is a two-layer fully connected neural network. σis the Sigmoid activation function.
f
n×n
is the convolution operation with a convolution kernel size of n×n.
4.1.3. Impro-ResNet50 model. In this paper, the cotton seed detection model is based on
the original ResNet50 network structure, but adds the CBAM attention mechanism after the
Stage2 residual module and redesigned the fully connected layer and classification output
layer. The Impro-ResNet50 model is shown in Fig 5.
The description of the cotton seed image detection procedure by Impro-ResNet50 is shown
in Fig 5. The first step was converting a cotton seed input image to 224×224×3 pixels through
pre-processing operations such as data enhancement and input into Impro-ResNet50. The
residual block was then used to extract high-level characteristics from the image of cotton
seed. By assigning greater weights to the most significant feature channels and smaller weights
to the less substantial feature channels, the CBAM module was used to optimize the parts. As a
consequence of the pre-convolution operation, the Impro-ResNet50 model would not lose any
additional crucial information about cotton seeds due to the increased global attention. Finally,
Fig 4. CBAM module.
https://doi.org/10.1371/journal.pone.0273057.g004
Fig 5. The improved ResNet50 model.
https://doi.org/10.1371/journal.pone.0273057.g005
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 7 / 19
different classes of cotton seeds can be distinguished using the classifier pair.
4.2 Network training strategies
4.2.1. Transfer learning. Transfer learning is applying knowledge learned in one source
domain to another related target domain. Annotating large amounts of data in convolutional
neural networks can be prevented, the model’s dependence on data can be reduced, and the
training efficiency of the model can be enhanced [35–37]. This study was motivated by this
and trained the Impro-ResNet50 model using transfer learning.
ResNet50 was initially pre-trained on the massive public dataset ImageNet to obtain an ini-
tial converged weight in this study. This weight was then transferred to the Impro-ResNet50
model, which was trained using the previously self-constructed cotton seed dataset to generate
new weights. Finally, the parameters of the Impro-ResNet50 model were fine-tuned to
improve the model’s learning performance for this dataset. Using transfer learning for weight
initialization instead of random initialization of weights could accelerate the model’s conver-
gence and enhance its generalization capability.
4.2.2. Activation function. The Relu activation function is widely utilized in CNN due to
its quick operation and high performance. However, when the input was less than zero, the
Relu activation could not continue to update the neuron death parameters. In the LRelu activa-
tion function, the activation value was determined by a threshold, and the parameters could
continue to be updated if the input was less than 0. Although it addressed the issue of neural
death, the LReLu function was not as smooth as the ReLu function. The Softplus activation
function avoided the drawback of the Relu activation function’s forced sparsity. Similar to the
Relu function, it failed to address the function output offset phenomenon, negatively impact-
ing the model’s convergence performance [38–40]. The LRelu-Softplus activation function
was designed by combining the characteristics of the three activation functions under appeal.
The calculations of the four activation functions are given in Eqs 4to 7.
fðxÞ ¼ 0;x0
x;x>0ð4Þ
(
where x 0, the output is 0, and the neuron is inactivated.
fðxÞ ¼ ax;x0
x;x>0ð5Þ
(
where α= 0.01, x<0, the output is negative, and the neuron is still active.
fðxÞ ¼ lnðexþ1Þ ð6Þ
fðxÞ ¼ ax;x0
lnðexþ1Þ ln2;x>0ð7Þ
(
where α= 0.15. The LReLu-Softplus activation function is shown in Fig 6.
4.2.3. Optimisation algorithm. The optimizer is to update the network weights during
the network training so that the model gets the optimal value. Adam is the most popular opti-
mizer. The Adam algorithm estimates each gradient component’s first and second-order
moments to obtain the updated amount at each step and provides an adaptive learning rate
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 8 / 19
[41,42]. The calculation of Adam’s algorithm is given in Eq 8.
gt¼ ryftðyt1Þ
^
mt¼b1mt1þ ð1b1Þ gt
1bt
1
^
vt¼b2vt1þ ð1b2Þ g2
t
1bt
2
yt¼yt1a^
mt
ffiffiffiffi
^
vt
pþε
ð8Þ
8
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
:
where tis the number of time steps, θ
t
is the update gradient. g
t
is the first-order derivative. β
1
,
β
2
2[0,1) is the exponential decay rate. m
t
is the estimate of the first-order moments, ^
mtis the
bias-corrected estimate of the first-order moments. v
t
is the estimate of the second-order
moments, ^
vtis the bias-corrected estimate of the second-order moments. αis the step size. εis
an arbitrarily small positive number.
5. Experimental setup and evaluation indicators
5.1. Training platform and parameter settings
This test platform’s software environment was a Windows 10 64-bit system with 16 GB of
RAM. The CPU was an Intel Xeon E7, and the GPU was an NVIDIA GTX 1060. Pytorch used
Python 3.8 as the programming language and Pytorch 1.9 as the deep learning framework to
implement parallel processing of convolutional neural networks on the GPU.
Fig 6. LRelu-Softplus activation function.
https://doi.org/10.1371/journal.pone.0273057.g006
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 9 / 19
The Adam optimization algorithm was chosen for the Impro-ResNet50 model with exponen-
tial decay rates of 0.9 and 0.999, respectively, and Eps of 1e−08. The convergence rate of the
model was determined by the learning rate, which was set to 1e-04 in this study. Taking into
account the training effect of the model and the experimental conditions, the batch size was set to
16, so 16 samples were entered into the model each time. To prevent overfitting, dropout was
implemented before the final layer of the model to deactivate neurons with a predetermined prob-
ability, reduce the dependence between neurons, and enhance the model’s ability to generalize.
The dropout value was set to 0.45. The value of Epochs was set to 300. The images are then nor-
malized before being fed into the CNN. The experiment’s hyperparameters are shown in Table 2.
5.2. Network training process
Transfer learning was utilized in the training of the Impro-ResNet50 model. The training pro-
cess of the model is shown in Fig 7. Initially, the cotton image dataset was loaded into the
Pytorch deep learning framework and divided into training and validation sets using the dataset
loading method. The ResNet50-pre.pth pre-training model should then be loaded. On the train-
ing set, the Impro-ResNet50 model was trained, and on the validation set, model evaluation
results were obtained for each number of iterations. The cross-loss entropy function produced a
gradual reduction in loss and increased precision. The model’s training was concluded after 300
iterations, and the best training model was saved. Using the newly collected, unlabeled images
of cotton seed, the best-trained model was identified, and the prediction results were presented.
5.3. Evaluation metrics
The confusion matrix’s calculated accuracy (A
cc
), precision (P
r
), recall (R
e
), and F1-score (F
1
)
were used as evaluation metrics in this study. Time spent processing a single image (T
s
) was
also crucial for evaluating models. Short training times for models are the solution to computa-
tional resource constraints [43–45]. The calculations of the five evaluation indicators are given
in Eqs 9to 13.
Acc ¼TpþTn
TpþFpþTnþFnð9Þ
Precision ¼Tp
TpþFpð10Þ
Recall ¼Tp
TpþFnð11Þ
Table 2. Hyperparameters of Impro-ResNet50 model.
Parameters Values
Optimizer Adam
Learning rate 1e−04
Betas (β
1
,β
2
) 0.9, 0.999
Eps(ε) 1e−08
Batch_size 16
Epochs 300
Dropout 0.45
Target_size 224×224×3
https://doi.org/10.1371/journal.pone.0273057.t002
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 10 / 19
where T
p
,T
n
,F
p
, and F
n
demonstrate the true positive, true negative, false positive, and false
negative, respectively. Tis total train time. Nis the total number of train images.
6. Results
In this section, a re-collected dataset containing 450 unannotated images of intact, broken and
cracked cotton seeds (150 of each type) was used for the performance evaluation of all models.
6.1. The impact of parameter optimization on the Impro-ResNet50 model
To determine the effects of the learning rate, activation function, and fully connected layer
design on the performance of the Impro-ResNet50 model, the following three controlled
experiments were designed. Experiment 1 compared the impact of various learning rates on
the model’s performance. Experiment 2 compared the impact of different activation functions
on model performance. Experiment 3 compared the impact of fully connected layer layouts on
model performance.
1. The impact of learning rate on the model.
The model converged slowly when Adam’s optimization algorithm’s learning rate was too
low. A setting that was too large leads to non-convergence, and the loss function misses the
optimal solution. The initial batch size was determined to be sixteen. In the Adam optimiza-
tion algorithm, the default learning rate value was 0.001. To compare the training effect of the
model, different orders of magnitude of parameter values were used, including 0.1, 0.01, 0.001,
0.0001 and 0.00001. The effect of learning rate adjustment on model loss values and accuracy
is shown in Fig 8.
As shown in Fig 8, the model converges slowly, at a learning rate of 0.1. At a learning rate of
0.00001, the model barely converges, and the loss value is significant. When the learning rate
was 0.01, 0.001, or 0.0001, the model converged well. However, when the learning rate was
0.0001, the model converged the quickest and had the smallest loss value after convergence.
The model was tested with the highest accuracy at a learning rate of 0.0001 after 300 rounds of
training. Consequently, the learning rate was set to 0.0001 during the Impro-ResNet50 model’s
training.
2. The impact of the activation function on the model.
Fig 8. Effects of learning rate on training accuracy and training loss.
https://doi.org/10.1371/journal.pone.0273057.g008
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 12 / 19
The activation function played a crucial role in training the CNN, which provided the
model with a robust capacity for fitting. The training effect of the model was compared when
it was trained with Relu, LRelu, Softplus, and LRelu-Softplus activation functions using the
fixed learning rate value of 0.0001. The effects of different activation functions on the loss val-
ues and accuracy of the model are shown in Fig 9.
As the number of iterations increased, the loss value of the model decreased, and the LRelu-
Softplus activation function produced the fastest convergence and most stable training results,
as shown in Fig 9. The other three activation functions exhibited more significant fluctuations
in the training curve and larger loss values after the training was completed. Using the LRelu-
Softplus activation function to train the model increased robustness and precision. The
Impro-ResNet50 model was therefore trained using the LRelu-Softplus activation function.
3. Impact of fully connected layers on the model
In CNN, the fully connected layer combines the feature and classifier functions. Fully-con-
nected layers contained a large number of model-size-affecting parameters. To determine the
optimal number of fully connected model layers, the impact of adding one to three fully con-
nected layers on model performance was compared. The effect of a different number of fully
connected layers on the loss value and accuracy of the model is shown in Fig 10.
The model performs worst when configured with three fully connected layers, as shown in
Fig 10. This could be due to the fact that the three-layer fully-connected layer resulted in exces-
sive model parameters and overfitting during training. When comparing the model with one
and two fully-connected layers, the model trained with two fully-connected layers had a
smoother convergence and a lower loss value at the end of training. The Impro-ResNet50
model was therefore trained with two fully connected layers.
6.2. Effect of attention mechanism on model performance
To further validate the benefits of the model incorporating the CBAM attention mechanism,
the CBAM attention mechanism was substituted with the ultra-lightweight attention mecha-
nism models SE and CA for comparison experiments conducted under the same experimental
conditions [46–49]. The experimental schemes were as described below. Scheme 1 model with
no additional attention mechanism. Scheme 2 replaced the CBAM module for the SE module.
Scheme 3 replaced the CBAM module for the CA module. Scheme 4 was the Impro-ResNet50
model of this paper. The results of the three experimental schemes are shown in Table 3.
Fig 9. Effects of activation function on training accuracy and training loss.
https://doi.org/10.1371/journal.pone.0273057.g009
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 13 / 19
As shown in Table 3, the Impro-ResNet50 model achieved an average detection accuracy of
97.23% for cotton seeds, which was 0.82%, 1.11%, and 1.62% higher than the other three exper-
imental models. The time required to detect a single image was 0.11s, which was 0.04s, 0.08s,
and 0.38s faster than the other three experimental models. Comprehensive appeal results dem-
onstrated the efficacy of introducing an attention mechanism to improve the model’s accuracy.
Moreover, only the SE model utilized the channel attention mechanism. In contrast, the other
attention models presented in the paper were an organic combination of the channel attention
mechanism and location feature data. Experiments comparing SE, CA, and CBAM attention
revealed that spatial feature information contributed to the model’s enhanced performance. In
particular, the CBAM module improved model accuracy to the greatest extent, and the embed-
ding of the CBAM module into the ResNet50 model enabled the model to simultaneously
acquire channel information as well as spatial information about cotton seed regions, thereby
improving the model’s learning ability.
6.3. Performance comparison of other models
To further demonstrate the detection ability of the Impro-ResNet50 model, AlexNet, VGG16,
GoogLeNet (InceptionV3), EfficientNet, and ResNet18 were chosen for migration learning
and compared under identical experimental conditions. The experimental outcomes are
shown in Table 4.
As shown in Table 4, the Impro-ResNet50 model outperformed the other five classical
models in terms of detection accuracy and training time. The Impro-ResNet50 model detected
cotton seeds with an average accuracy of 97.23%, which was 1.69%, 2.21%, 2.39%, 3.16%, and
5.05% higher than a variety of other models. In addition, the processing time for a single
image was 0.11s, which was the fastest among all models. In the meantime, recall, precision,
Fig 10. Effects of fully connected layers on training accuracy and training loss.
https://doi.org/10.1371/journal.pone.0273057.g010
Table 3. Comparison of the results of cotton seed detection with different attention mechanisms.
No. Model names P
r
/% R
e
/% F
1
/% Params / M T
s
/s A
cc
/%
1 ResNet50 95.55 95.62 95.58 30.2 0.49 95.61
2 Impro-ResNet50-SE 96.00 96.05 96.02 31.5 0.19 96.12
3 Impro-ResNet50-CA 96.45 96.27 96.36 32.4 0.15 96.41
4 Impro-ResNet50 97.33 97.13 97.22 33.4 0.11 97.23
https://doi.org/10.1371/journal.pone.0273057.t003
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 14 / 19
and F1 all achieved positive outcomes. Although the number of parameters was not the lowest,
it was within the affordability range for hardware. In terms of overall performance, the advan-
tages of the model proposed in this paper were greater. The AlexNet model required the most
time and had the lowest detection accuracy among the five classical models. The VGG16
model had the most parameters and required a substantial amount of computational
resources. The GoogLeNet model implemented the Inception structure, which drastically
reduced the number of model parameters and improved detection performance. EfficientNet
utilized NAS technology to simultaneously search and optimize the model depth, width, and
input image resolution in order to extend the model structure proportionally and attain a high
level of structural proportionality. Therefore, the detection task also yielded good results. The
ResNet18 model had a similar structure to the Impro-ResNet50 model, utilizing the residual
structure to enhance its feature learning capability. However, its residual block consisted of 18
layers. Although the parameters were reduced compared to the Impro-ResNet50 model, the
time consumption and detection accuracy in a single image were also diminished. The
improved Impro-ResNet50 model could detect images of cotton seeds with greater precision.
6.4. Confusion matrix to visualize and analyse model detection results
A confusion matrix is a valuable tool for evaluating the quality of a classification model and its
performance. Each row represents the actual data for a category, while each column represents
the predicted data for that category, with the diagonal values indicating the likelihood of being
accurate. The confusion matrix of the Impro-ResNet50 model, as shown in Fig 11.
As shown in Fig 11, the average classification accuracy of the model was 97.23%, and the
classification performance (in terms of F1 score) for broken, intact, and cracked cotton seeds
decreased from highest to lowest. There were 147 correct identifications out of 150 intact cot-
ton seeds, 146 correct identifications out of 150 broken cotton seeds, and 145 correct identifi-
cations out of 150 cracked cotton seeds. By analyzing the misclassified images, it was
determined that the Impro-ResNet50 model had a high misclassification rate when classifying
cracked cotton seeds as intact cotton seeds. It was difficult for the model to detect cracked fea-
tures in the images because they were dark, the overall resolution was low, and factors such as
the angle of the shot made it difficult for the model to detect them.
7. Conclusion and future work
The construction of an attention-based mechanism for the cotton seed quality detection
model. Integrating feature channels and spatial location information was accomplished by
incorporating a CBAM module. A modified LRelu-Softplus activation function was used to
enhance the model’s capacity for generalization. The transfer learning strategy and Adam opti-
mization training algorithm decreased model parameters and accelerated model convergence
speed. The influence of parameter settings and attention mechanisms on the model was
Table 4. Comparison of test results for different model cotton seeds.
Model names P
r
/% R
e
/% F
1
/% Params / MT
s
/s A
cc
/%
AlexNet 92.00 92.21 92.10 62.7 1.02 92.18
VGG16 94.00 94.09 94.04 145.2 0.87 94.07
GoogLeNet 94.44 94.52 94.48 24.6 0.62 94.84
EfficientNet 95.11 94.99 95.02 42.6 0.21 95.02
ResNet18 95.33 95.39 95.35 15.2 0.18 95.54
Impro-ResNet50 97.33 97.13 97.22 33.4 0.11 97.23
https://doi.org/10.1371/journal.pone.0273057.t004
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 15 / 19
discussed and compared with AlexNet, VGG16, GoogLeNet, EfficientNet, and ResNet18. The
following are the conclusions:
1. The Impro-ResNet50 model constructed with a learning rate of 0.0001, an activation func-
tion of LRelu-Softplus, and two fully connected layers converged the quickest and were the
most robust. After training, the average detection accuracy of the Impro-ResNet50 model
reached 97.23%, and the time required to process a single image was only 0.11s.
2. Compared to the three models without an embedded attention mechanism, the embedded
SE attention mechanism, and the embedded CA attention mechanism, the average detec-
tion accuracy was improved by 0.82%, 1.11%, and 1.62%, respectively. The processing time
of a single image was enhanced by 0.04s, 0.08s, and 0.38s, respectively, under identical
experimental conditions.
3. Compared to traditional models such as AlexNet, VGG16, GoogleNet, EfficientNet, and
ResNet18, the average detection accuracy was increased by 1.69–5.05% and the time
required to process a single image was decreased by 0.07–0.91s.
4. The confusion matrix revealed that the Impro-ResNet50 model had a higher overall recog-
nition accuracy and produced superior results for cotton seeds. However, the model still
has a certain misclassification rate, with the detection of cracked cotton seeds performing
the worst. Future research will be conducted on detecting cracked cotton seeds with less
obvious taxonomic features.
The Impro-ResNet50 cotton seed quality detection model, based on the attention mecha-
nism, was trained on a large amount of data while maintaining high accuracy and requiring
Fig 11. Confusion matrix for the Impro-ResNet50 model.
https://doi.org/10.1371/journal.pone.0273057.g011
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 16 / 19
only a short amount of time to run. In the future, we will supplement the data with cotton
seeds of different qualities and backgrounds so that the model has a wider range of applica-
tions. At the same time, the model is simplified so that it can be deployed on mobile and easily
used by farmers.
Supporting information
S1 Data.
(ZIP)
Author Contributions
Conceptualization: Xinwu Du.
Data curation: Pengfei Li, Zhihao Yun.
Funding acquisition: Xinwu Du.
Investigation: Laiqiang Si.
Methodology: Xinwu Du, Laiqiang Si, Pengfei Li, Zhihao Yun.
Project administration: Xinwu Du, Laiqiang Si, Pengfei Li, Zhihao Yun.
Resources: Xinwu Du, Laiqiang Si.
Software: Xinwu Du, Laiqiang Si.
Supervision: Xinwu Du.
Validation: Xinwu Du.
Visualization: Xinwu Du.
Writing – original draft: Laiqiang Si.
Writing – review & editing: Xinwu Du.
References
1. Gao P, Zhang C, Lu X, Zhang Z, He Y. Visual identification of Slight-Damaged cotton seeds based on
near infrared hyperspectral imaging. Spectroscopy and Spectral Analysis. 2018; 38: 1712–1718.
2. Zhu S, Zhou L, Gao P, Bao Y, He Y, Feng L. Near-infrared hyperspectral imaging combined with deep
learning to identify cotton seed varieties. Molecules. 2019; 24: 3268. https://doi.org/10.3390/
molecules24183268 PMID: 31500333
3. Duan L, Yan T, Wang J, Ye W, Chen W, Gao P, et al. Combine hyperspectral imaging and machine
learning to identify the age of cotton seeds. Spectroscopy and Spectral Analysis. 2021; 41: 3857–3863.
4. Zhang J, Geng Y, Guo F, Li X, Wan S. Research progress on the mechanism of improving peanut yield
by single-seed precision sowing. Journal of Integrative Agriculture. 2020; 19: 1919–1927.
5. Zhao X, Chen L, Gao Y, Yang S, Zhai C. Optimization method for accurate positioning seeding based
on sowing decision. International Journal of Agricultural and Biological Engineering. 2021; 14: 171–
180.
6. Bao F, Bambil D. Applicability of computer vision in seed identification: deep learning, random forest,
and support vector machine classification algorithms. Acta Botanica Brasilica. 2021; 35: 17–21.
7. Koklu M, Ozkan IA. Multiclass classification of dry beans using computer vision and machine learning
techniques. Computers and Electronics in Agriculture. 2020; 174: 105507.
8. Jin B, Qi H, Jia L, Tang Q, Gao L, Li Z, et al. Determination of viability and vigor of naturally-aged rice
seeds using hyperspectral imaging with machine learning. Infrared Physics &Technology. 2022; 122:
104097.
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 17 / 19
9. Przybył K, Wawrzyniak J, Koszela K, Martynenko A. Application of deep and machine learning using
image analysis to detect fungal contamination of rapeseed. Sensors. 2020; 20: 7305. https://doi.org/
10.3390/s20247305 PMID: 33352649
10. Javanmardi S, Ashtiani SHM, Verbeek FJ, et al. Computer-vision classification of corn seed varieties
using deep convolutional neural network. Journal of Stored Products Research. 2021; 92: 101800.
11. Singh T, Garg NM, Iyengar SRS. Nondestructive identification of barley seeds variety using near-infra-
red hyperspectral imaging coupled with convolutional neural network. Journal of Food Process Engi-
neering. 2021; 44: e13821.
12. Mukasa P, Wakholi C, Faqeerzada MA, Amanah HZ, Kim H, Joshi R, et al. Nondestructive discrimina-
tion of seedless from seeded watermelon seeds by using multivariate and deep learning image analysis.
Computers and Electronics in Agriculture. 2022; 194: 106799.
13. Sabanci K, Aslan MF, Ropelewska E, Unlersen MF. A convolutional neural network-based comparative
study for pepper seed classification: Analysis of selected deep features with support vector machine.
Journal of Food Process Engineering. 2022; 45: e13955.
14. Weng S, Han K, Chu Z, Zhu G, Liu C, Zhu Z, et al. Reflectance images of effective wavelengths from
hyperspectral imaging for identification of Fusarium head blight-infected wheat kernels combined with a
residual attention convolution neural network. Computers and Electronics in Agriculture. 2021; 190:
106483.
15. Huang S, Fan X, Sun L, Shen Y, Suo X. Research on classification method of maize seed defect based
on machine vision. Journal of Sensors. 2019; 2019: 1–9.
16. Quan L, Zhang T, Sun L, Chen X, Xu Z. Design and testing of an on-line omnidirectional inspection and
sorting system for soybean seeds. Applied Engineering in Agriculture. 2018; 34: 1003–1016.
17. Tu K, Li L, Yang L, Wang J, Sun Q. Selection for high quality pepper seeds by machine vision and clas-
sifiers. Journal of Integrative Agriculture. 2018; 17: 1999–2006.
18. Lin P, Li X, Li D, Jiang S, Zou Z, Lu Q, et al. Rapidly and exactly determining postharvest dry soybean
seed quality based on machine vision technology. Scientific Reports. 2019; 9: 1–11.
19. Salimi Z, Boelt B. Classification of processing damage in sugar beet (Beta vulgaris) seeds by multispec-
tral image analysis. Sensors. 2019; 19: 2360. https://doi.org/10.3390/s19102360 PMID: 31121960
20. Yasmin J, Lohumi S, Ahmed MR, Kandpal LM, Faqeerzada MA, Kim MS, et al. Improvement in purity of
healthy tomato seeds using an image-based one-class classification method. Sensors. 2020; 20: 2690.
https://doi.org/10.3390/s20092690 PMID: 32397311
21. de Medeiros AD, Pinheiro DT, Xavier WA, da Silva LJ, Dias DCFD. Quality classification of Jatropha
curcas seeds using radiographic images and machine learning. Industrial Crops and Products. 2020;
146: 112162.
22. Meng F, Luo S, Sun H, Li M. Design and Experiment of Real-time Detection and Sorting Device for
Maize Seeds. Transactions of the Chinese Society for Agricultural Machinery. 2021; 52: 153–159,177.
23. Zhang J, Qu M, Gong Z, Cheng F. Online double-sided identification and eliminating system of
unclosed-glumes rice seed based on machine vision. Measurement. 2022; 187: 110252.
24. AltuntaşY, Co
¨mert Z, Kocamaz AF. Identification of haploid and diploid maize seeds using convolu-
tional neural networks and a transfer learning approach. Computers and Electronics in Agriculture.
2019; 163: 104874.
25. Zhang S, Zhang Q, Li K. Detection of peanut kernel quality based on machine vision and adaptive con-
volution neural network. Transactions of the CSAE. 2020; 36: 269–277.
26. Ma T, Tsuchikawa S, Inagaki T. Rapid and non-destructive seed viability prediction using near-infrared
hyperspectral imaging coupled with a deep learning approach. Computers and Electronics in Agricul-
ture. 2020; 177: 105683.
27. Zhao G, Quan L, Li H, Feng H, Li S, Zhang S, et al. Real-time recognition system of soybean seed full-
surface defects based on deep learning. Computers and Electronics in Agriculture. 2021; 187: 106230.
28. Thakur PS, Tiwari B, Kumar A, Gedam B, Bhatia V, Krejcar O, et al. Deep transfer learning based pho-
tonics sensor for assessment of seed-quality. Computers and Electronics in Agriculture. 2022; 196:
106891.
29. Hao L, Liu Z, Sun H, Rao Z, Ji H. Discrimination of unsound wheat kernels based on deep convolutional
generative adversarial network and near-infrared hyperspectral imaging technology. Spectrochimica
Acta Part A: Molecular and Biomolecular Spectroscopy. 2022; 268: 120722. https://doi.org/10.1016/j.
saa.2021.120722 PMID: 34902690
30. Wang L, Liu J, Zhang J, Wang J, Fan X. Corn Seed Defect Detection Based on Watershed Algorithm
and Two-Pathway Convolutional Neural Networks. Frontiers in Plant Science. 2022; 13:730190.
https://doi.org/10.3389/fpls.2022.730190 PMID: 35283875
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 18 / 19
31. Sreedhar PSS, Satya S, Nandhagopal N. Classification Similarity Network Model for Image Fusion
Using Resnet50 and GoogLeNet. Intelligent automation and soft computing. 2022; 31: 1331–1344.
32. Xu W, Yan Z. Research on strawberry disease diagnosis based on improved residual network recogni-
tion model. Mathematical Problems in Engineering. 2022; 2022: 1–13.
33. Zhang W, Ma H, Li X, Liu X, Jiao J, Zhang P, et al. Imperfect Wheat Grain Recognition Combined with
an Attention Mech-anism and Residual Network. Applied sciences. 2021; 11: 5139.
34. Zhao Y, Sun C, Xu X, Chen J. RIC-Net: A plant disease classification model based on the fusion of
Inception and residual structure and embedded attention mechanism. Computers and Electronics in
Agriculture. 2022; 193: 106644.
35. Zhao X, Li K, Li Y, Ma J, Zhang L. Identification method of vegetable diseases based on transfer learn-
ing and attention mechanism. Computers and Electronics in Agriculture. 2022; 193: 106703.
36. Lai M, Gao L. Automatic classification of apple leaf diseases based on transfer learning. International
journal of Robotics & Automation. 2022; 37: 44–51.
37. Wu N, Liu F, Meng F, Li M, Zhang C, He Y. Rapid and accurate varieties classification of different crop
seeds under sample-limited condition based on hyperspectral imaging and deep transfer learning. Fron-
tiers in Bioengineering and Biotechnology. 2021; 9:696292. https://doi.org/10.3389/fbioe.2021.696292
PMID: 34368096
38. Varshney M, Singh P. Optimizing nonlinear activation function for convolutional neural networks. Signal,
Image and Video Processing. 2021; 15: 1323–1330.
39. Swiderski B, Osowski S, Gwardys G, et al. Random CNN structure: tool to increase generalization abil-
ity in deep learning. Eurasip journal on image and video processing. 2022; 2022: 1–12.
40. Kilic¸arslan S, Celik M. KAF+ RSigELU: a nonlinear and kernel-based activation function for deep neural
networks. Neural Computing and Applications. 2022; 2022: 1–15.
41. Waheed A, Goyal M, Gupta D, Khanna A, Hassanien AE, Pandey HM, et al. An optimized dense convo-
lutional neural network model for disease recognition and classification in corn leaf. Computers and
Electronics in Agriculture. 2020; 175: 105456.
42. Nanni L, ManfèA, Maguolo G, et al. High performing ensemble of convolutional neural networks for
insect pest image detection. Ecological Informatics. 2022; 67: 101515.
43. Hussain N, Farooque AA, Schumann AW, Abbas A, Acharya B, McKenzie-Gopsill A, et al. Application
of deep learning to detect Lamb’s quarters (Chenopodium album L.) in potato fields of Atlantic Canada.
Computers and Electronics in Agriculture. 2021, 182: 106040.
44. Moses K, Miglani A, Kankar PK. Deep CNN-based damage classification of milled rice grains using a
high-magnification image dataset. Computers and Electronics in Agriculture. 2022; 195: 106811.
45. Zhao S, Liu J, Bai Z, et al. Crop pest recognition in real agricultural environment using convolutional
neural networks by a parallel attention mechanism. Frontiers in Plant Science. 2022; 13:839572.
https://doi.org/10.3389/fpls.2022.839572 PMID: 35265096
46. Sun X, Ma L, Li G. Multi-vision attention networks for on-line red jujube grading. Chinese Journal of
Electronics. 2019; 28: 1108–1117.
47. Zhu X, Zhang X, Sun Z, Zheng Y, Su S, Chen F. Identification of oil tea (camellia oleifera c. abel) culti-
vars using EfficientNet-B4 CNN model with attention mechanism. Forests. 2021; 13: 1.
48. Wang P, Niu T, Mao Y, Zhang Z, Liu B, He D, et al. Identification of apple leaf diseases by improved
deep convolutional neural networks with an attention mechanism. Frontiers in Plant Science. 2021; 12:
1997. https://doi.org/10.3389/fpls.2021.723294 PMID: 34650580
49. Bhujel A, Kim NE, Arulmozhi E, Basak JK, Kim. A lightweight attention-based convolutional neural net-
works for tomato leaf disease classification. Agriculture. 2022; 12: 228.
PLOS ONE
Cotton seed quality detection
PLOS ONE | https://doi.org/10.1371/journal.pone.0273057 February 15, 2023 19 / 19
