www.nature.

com/scientificreports

OPEN Innovative adaptive edge detection

for noisy images using wavelet and
Gaussian method
Huanxu Li1,2 & Keke Xu1,2
Edge detection is a crucial task in image processing and remote sensing, particularly for accurately
identifying and separating shapes in noisy digital images. To enhance robustness and detail in edge
detection, this study presents an innovative edge detection method, which integrates a denoising
module and an adaptive thresholding technique to effectively address challenges associated with
Gaussian noise in images. The proposed denoising module employs wavelet and Gaussian denoising
functions to decompose, filter, and reconstruct the image, thereby reducing the impact of noise and
enhancing image quality. For edge detection, an adaptive thresholding method based on a modified
OTSU method is utilized. Comprehensive experiments validate the proposed framework by comparing
detected edges against ground truth across various levels of Gaussian noise (0.1%, 10%, 20%, and
30%). The median thresholding function is chosen for its stability and convenience, while hard
thresholding is avoided due to its tendency to introduce artifacts. Objective metrics, including Mean
Squared Error (MSE), Accuracy, and Peak Signal-to-Noise Ratio (PSNR), are employed for evaluation.
Comparative results indicate that the proposed method outperforms traditional methods, such as
Canny and Roberts, showcasing its effectiveness in edge detection.

Keywords Edge detection, Wavelet transform, Wavelet threshold denoising, Maximum inter-class variance
method (OTSU), Adaptive threshold

Edge detection is a fundamental technique in remote sensing, image processing and computer vision, and its
performance directly affects the final results that are produced by various applications, especially in scenarios
with high precision requirements, such as, segmentation1–3, interest point extraction4,5, image recognition6,7,
object tracking8–13, image retrieval14,15, etc. In scenarios requiring high precision, the effectiveness of edge
detection algorithms is crucial and extensively studied.
However, image denoising and fixed thresholding can result in linear translation or rotation of edge
contours16,17. Traditional edge detection techniques, like the Robert18, Prewitt19, Canny20,21, and Sobel22
operators, have limitations in producing sharp edges and are sensitive to image noise, resulting in the decrease
of detection accuracy. To overcome the issue, the existing methods mainly leverage simple linear translation-
invariant filters to preprocess images, such as the mean23, Gaussian24, and Laplacian25. However, the linear filters
may blur and weaken the contrast across edges, or even worse, blend adjacent edges26. In the field of image
denoising, deep learning methods have made significant advancements in recent years. Generative adversarial
network (GAN) combined with a semi-soft thresholding model27 is used to enhance image quality by reducing
noise while preserving edges and details. Pulse coupled neural networks (PCNN)28 for noise detection with
wavelet filtering and Perona-Malik regularization is used to effectively remove mixed noise while maintaining
image integrity. Image denoising affects the quality of subsequent processing steps; therefore, it is necessary to
apply appropriate filtering (denoising) processing to the image so as to eliminate the noise and retain the original
edge information to the maximum extent.
On the other hand, using fixed thresholding methods for edge detection can lead to suboptimal performance,
potentially missing or misidentifying edges. Manual adjustments to the threshold may be necessary when
background or illumination conditions change. Adaptive thresholding techniques, like the statistic information
of the gradient histogram29–31, OTSU32–37, 2D image entropy38, etc., offer a more flexible approach to edge
detection.
Similarly, in edge detection, deep learning methods have shown remarkable capabilities. Many modern
edge detection algorithms, such as dense extreme inception network39, Tiny and Efficient Edge Detector40
and R-DbCNN41 to automatically learn edge features, significantly outperforming traditional techniques in

1School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo 454000, China.
2Huanxu Li and Keke Xu contributed equally to this work. email: [email protected]

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 1

www.nature.com/scientificreports/

complex scenarios. Despite the excellent performance of deep learning methods, traditional techniques still
hold advantages in terms of computational efficiency and interpretability in certain applications.
Traditional methods, such as the Canny edge detector and Sobel operator, often exhibit lower computational
complexity, particularly in resource-constrained environments like edge computing devices and mobile
applications, allowing for real-time processing without sacrificing performance42. Additionally, traditional
methods are more interpretable and provide clear decision-making rationale, which is especially important in
fields like medical image processing where users need to understand the workings of the algorithm to ensure the
reliability and effectiveness of its result43. These advantages affirm the continued practical relevance of traditional
methods in specific application scenarios. Therefore, we choose traditional methods for image edge detection.
To tackle critical challenges in edge detection, particularly in noisy environments, we introduce an innovative
edge detection method through adaptive thresholding and wavelet-Gaussian denoising (EDAW). This framework
specifically aims to reduce noise interference while preserving essential edge details. By systematically applying
the denoising process, EDAW ensures enhanced edge clarity. Following this, we implement gradient calculation
and non-maximal suppression (NMS) to accurately identify edge pixels. The edges are further refined through
adaptive thresholding and connection strategies, ensuring robustness against noise artifacts.
This paper presents a focused contribution:

1. Innovative Denoising Technique: EDAW introduces a novel denoising method that effectively reduces noise
while maintaining edge integrity, addressing common pitfalls associated with traditional denoising methods,
such as blurring and over-smoothing.
2. Adaptive Thresholding Method: We employ a refined adaptive thresholding method, based on the modified
OTSU technique, that enhances edge detection performance in the presence of noise without sacrificing
critical image details.
3. Comprehensive Performance Assessment: We rigorously evaluate EDAW’s performance against contempo-
rary edge detection methodologies using objective metrics such as Mean Squared Error (MSE), Accuracy,
and Peak Signal-to-Noise Ratio (PSNR). Our findings demonstrate EDAW’s superior robustness and preci-
sion, outperforming conventional methods like Canny across diverse levels of Gaussian noise.The remainder
of this paper is organized as follows. Section 2 represents the related works. Research methodology in detail
is presented in Section 3. Performance metrics along with result analyses are described in Section 4. Finally,
this paper reaches the conclusion in Section 5.

Related works
Basic principles of wavelet threshold denoising
Wavelet Threshold Denoising (WTD) is a multi-scale signal analysis and denoising method that involves a
reduced-order decomposition of the signal44,45.
After performing decomposition, a set of wavelet decomposition coefficients is obtained, and a threshold
is established for quantitative processing. The signal data undergo both low-pass and high-pass filtering. The
coefficients below the threshold, referred to as the low-pass filter, are separated into noise information and
set to zero. This portion primarily captures the high-frequency information of the signal through the detail
component. On the other hand, wavelet coefficients exceeding the threshold, known as the high-pass filter,
primarily represent the low-frequency information of the signal through the approximate component. It is then
determined whether the effective signal should be retained or subjected to further processing. Ultimately, the
denoised signal is achieved by reconstructing the processed wavelet coefficients, specifically the approximate
component.
Upon completing the wavelet reconstruction on the threshold detail components, we are able to acquire more
refined image information.
The flow chart of WTD is shown as Fig. 1.
Threshold functions, such as hard, soft, mean, and median, are employed to com-press wavelet coefficients.
While the hard threshold approach can introduce new flexible and unstable discontinuities with data variations,
soft thresholding may overly smooth the signal, leading to the loss of critical details and edges in the data.
Therefore, the median threshold function is chosen for its stability and effectiveness.

Image gradient calculation

After image denoising, the gradient and magnitude calculation are applied. Let the first order partial derivative’s
approximation on the X and Y directions be GX, GY respectively. The magnitude G can be calculated by:
√
G= (G2X + G2Y )(1)

where the X and Y are the directions of the image.

The direction theta can be calculated by:

θ = arctan(GY /GX )(2)

The temple of the Sobel operator on the X and Y direction is:

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 2

www.nature.com/scientificreports/

Fig. 1. The diagram of WTD.

[ ]
−1 0 1
SX = −2 0 2 (3)
−1 0 1
[ ]
1 2 1
SY = 0 0 0 (4)
−1 −2 −1

SY is the transpose matrix of SX . GX and GY can be obtained by the product of the equation ((3),(4)) and the
detection window. The magnitude G and the direction theta can be calculated by equation (1) and (2) respectively.

Non-maximum suppression (NMS)

After obtaining the gradient magnitude image G, it is essential to conduct non-maximum suppression (NMS) to
precisely locate edges. NMS guarantees that each edge is depicted as a single-pixel wide feature.
Typically, a 3 × 3 neighborhood, encompassing eight directions, is used to interpolate the gradient magnitude
in the direction of the gradient. If the magnitude G surpasses the values interpolated along the gradient direction,
the point is identified as a potential edge point; otherwise, it is categorized as a non-edge point. As a result, the
candidate edge image is produced through this procedure.

Adaptive threshold using OTSU

The OTSU method is used to find an adaptive threshold by maximizing interclass variance of grayscale values.
The principle of image processing is as follows:
Let N be the total number of pixels, the gradient magnitude be [1, L], Ni be the number of pixels with
gradient i. The gray-level histogram is normalized and the probability distribution is regarded as the number
pixel i:

∑
L

pi = Ni /N, pi ≥ 0, pi = 1(5)
i=1

The threshold T divide the gradients to two categories, i.e. [1, T] and [T+1, L]. The probabilities of two classes,
pb(T) and po(T), are given by equations (6) and (7).
( T )
∑
pb (T ) = ipi /wb (T ) (6)
i=1
( )
∑
L−1

po (T ) = ipi /wo (T ) (7)

i=T +1

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 3

www.nature.com/scientificreports/

∑T ∑L−1
where wb (T ) = i=1
pi , wo (T ) = T +1
pi , pb (T ) + po (T ) = 1. wb (T ) and wo (T ) are the number of
pixels in the corresponding area.

The total average gradient is defined by equation (8).

u = pb (T )wb (T ) + po (T )wo (T )(8)

The between-class variance for the two classes is defined by equation (9).

σ(T )2 = wb (T )[pb (T ) − u]2 + wo (T )[po (T ) − u]2 (9)

The optimal threshold T will maximize the equation.

Proposed method
Our purpose is to get the edge with robustness and high precision detection. The EDAW method we proposed
is shown in Fig. 2.

The denoising module

To simplify processing and focus on structural features, we convert the input color image to grayscale before
denoising. This step reduces computational load and color information.
In the denoising module, we enhance robustness by combining wavelet transform with a Gaussian filter
to eliminate noise. When using Gaussian filters for denoising, selecting a small sigma preserves more details
and edges in the image but may not effectively remove noise. Conversely, choosing a large sigma effectively
eliminates noise but can lead to blurring of details and edges. To maintain edge information, we opt for a small
sigma to initially smooth the input image.
To enhance the denoising effect, we incorporate wavelet transform as a secondary denoising method.
Wavelet-based methods are preferred for preserving edges and fine details compared to Gaussian denoising,
making them crucial for applications where maintaining image integrity is vital. Additionally, wavelets offer
spatial localization, allowing them to target noise in specific image regions effectively. The pseudo code is shown
in Fig. 3.

The edge extraction module

In the edge extraction module, the numerical gradient of the image is obtained and NMS is applied to suppress
non-edge pixels and reduce the width of the edges by 1 pixel.
The modified OTSU method to calculate the low and high thresholds is as follows.
The high and low threshold, Tl and Th, divide the gradient into three categories: [1, Tl], [Tl+1, Th], and
[Th+1, L] respectively. The average gradient values of the three categories are calculated by:
( Tl )
∑
p1 (T ) = ipi /w1 (T ) (10)
i=1
( )
∑
Th

p2 (T ) = ipi /w2 (T ) (11)

i=T l+1
( )
∑
L−1

p3 (T ) = ipi /w3 (T ) (12)

i=T h+1

Fig. 2. The flow chart of the EDAW method.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 4

www.nature.com/scientificreports/

Fig. 3. The pseudo code of denoising module.

∑T l ∑T h ∑L−1
where w1 (T ) = i=1
pi , w2 (T ) = T l+1
pi , w3 (T ) = T h+1
pi , p1 (T ) + p2 (T ) + p3 (T ) = 1. w1 (T ),
w2 (T ) and w3 (T ) are the number of pixels in the corresponding area.

The overall average gradient of the image is defined in equation (13).

u = p1 (T )w1 (T ) + p2 (T )w2 (T ) + p3 (T )w3 (T )(13)

The inter-class variances of the two categories are defined in equation (14).

σ(T )2 = w1 (T )[p1 (T ) − u]2 + w2 (T )[p2 (T ) − u]2 + w3 (T )[p3 (T ) − u]2 (14)

Then, all possible threshold combinations are explored to make the inter-class variance in formula (14) reaches
its maximum.

The low and high adaptive image threshold using OTSU are utilized to identify the edges. The pseudo code is
shown in Fig. 4. The pixels whose gradient magnitude is above the high-threshold will be marked as strong edge
points, and those whose gradient magnitude is under the low-threshold will be marked as non-edge points, and
the rest will be marked as weak candidate edge points. Lastly, those candidate edge pixels who are connect with
edge points will be marked as edge points. This connecting method reduces the influence of noise on the edge
of the final edge image.
The workflow of the EDAW algorithm is shown in Table 1.

The computational complexity

In this study, we implemented an image processing program based on the Canny edge detection algorithm.
The computational complexity of this algorithm is composed of several steps. Firstly, the complexity of the
noise addition step is O(N), where N represents the number of pixels in the image. Next, the wavelet denoising
operation typically has a complexity of O(N log N), depending on the implementation of the wavelet transform.
The steps for Gaussian denoising and calculating the image gradients also have a complexity of O(N), as do

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 5

www.nature.com/scientificreports/

Fig. 4. The pseudo code of modified OTSU method.

Input: Input image p, Gaussian standard deviation sigma

Output: Predicted edge I
Start
1. Convert the input image p to a grayscale image q.
2. Select WTD filtering function and decomposition level to smooth the grayscale image q1.
3. Smooth image q1 to image q2 by Gaussian filter of standard deviation sigma.
4. Calculate the gradient magnitude and direction, GX , GY by Sobel operator and theta.
5. Perform NMS in eight directions (E, NE, N, NW, W, SW, S, SE).
Gp1 = (1 − tan(θ)) × E + tan(θ) × N E
Gp2 = (1 − tan(θ)) × W + tan(θ) × SW
if GP ≥ Gp1 and GP ≥ Gp2
GP may be an edge.
else
GP should be suppressed.
6. Determine the high and low thresholds by the integral image. Select by:
if GP ≥ HighT hreshold
GP is a strong edge point.
else if GP ≥ LowT hreshold
GP is a weak edge point.
else
GP should be suppressed.
7. Select pixels between two thresholds and connect the edge pixels.
if Gp == LowT hreshold and Gp connected to a strong edge pixel
GP is a strong edge point.
else
GP should be suppressed.
8. Output Gp as the predicted edge I.
End

Table 1. The EDAW edge extraction algorithm.

the computations for gradient magnitude and angle, gradient quantization, non-maximum suppression, and
thresholding, all of which are O(N). However, the hysteresis thresholding step can reach a worst-case complexity
of O(N 2 ), as it may require multiple iterations to update edge pixels. Therefore, the overall worst-case
computational complexity of the algorithm is O(N 2 ), while in certain implementations, the average complexity
can be estimated as O(N log N). This complexity analysis provides important theoretical insights for subsequent
performance optimizations.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 6

www.nature.com/scientificreports/

Results
To achieve effective denoising, we applied ten different WTD methods, including bior4.4, bl7, coif5, db5, fk6,
haar, han5.5, mb4.2, rbio4.4, and sym546–53. The wavelet families we selected (such as bior4.4, db5, and coif5)
possess good approximation capability, orthogonality, and reconstruction ability, enabling effective noise
removal while preserving the signal characteristics. The wavelet decomposition level of three has been proven
to have a good denoising effect54, so we choose the wavelet decomposition level to be three. Denoising was per-
formed using a posterior median threshold function, with a Gaussian function sigma set at 0.8. Additionally, the
Canny algorithm was utilized in experiments for comparison purposes. The accuracy, peak signal-to-noise ratio
(PSNR), and mean squared error (MSE) of the reconstructed results were then evaluated.

Implemental details
The software environment of the experiment is windows 10 professional version. The “Butterfly” image processed
on a laptop with 12 GB RAM, NVidia GPU 3070. The Data processing, data graph display and generation are
processed by MATLAB R2024a. The code of deep learning is implemented with python and we used PyTorch
library.

Dataset
We used two datasets for our experiments: the “Butterfly” image and the publicly available dataset BSDS500.
The original image “Butterfly” is from the dataset55 in Kaggle. We stretched and blurred the details of the
butterfly torso from the original image to make it more conducive for edge detection. The test image “Butterfly”
and the ground truth image (GTI) are shown in Fig. 5a, b. To verify the robustness of our proposed method,
different levels of Gaussian noise are added to the original image (Fig. 5).
BSDS50056 is a dataset for edge detection and image segmentation released by Berkeley University. It contains
500 high-resolution natural images that have been manually annotated with edges and segmentation regions.
The images in the dataset typically have high resolution, making them suitable for evaluating and comparing
different image processing algorithms.

Criteria
To estimate the quality of the extracted, five metrics are used to compute the difference of the ground truth and
extracted edge image.

• Mean squared error (MSE)

Mean squared error (MSE) is estimate the similarity between the two images, given by:

M −1 N −1
1 ∑∑
M SE = [I(i, j) − K(i, j)]2 (15)
M ·N
i=0 j=0

where M, N is the total number of the GTI and extracted edge image, respectively.

Fig. 5. (a) Original image, (b) Ground truth image, (c) Image with 0.1% Gaussian noise, (d) Image with 10%
Gaussian noise, (e) Image with 20% Gaussian noise, (f) Image with 30% Gaussian noise.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 7

www.nature.com/scientificreports/

• Peak signal-to-noise ratio (PSNR) The peak signal-to-noise ratio (PSNR) is the ratio of the maximum possible
power of the information to the noise that affects its fidelity, given by:

P SN R = 20lg(M AX1 ) − 10lg(M SE)(16)

where MSE is the mean square error between images and MAX1 is the maximum possible pixel value in the
images.

• Accuracy

Accuracy is to Calculate the accuracy pixels between the GTI and extracted edge image, given by:

TP + TN
Accuracy = (17)
TP + TN + FP + FN

where TP is the true positive. TN is the true negative. FP is the false positive. FN is the false negative.

• Figure of Merit (FOM)

The Figure of Merit (FOM) is a quantitative measure used to evaluate the performance or efficiency. The
equation is given by:

NE
1 ∑ 1
F OM (E, G) = (18)
max(NE , NG ) 1 + αd2 (k))
k=1

where NG is the number of the actual edge points, NE is the number of the detected edge pixels, α is the scaling
constant, and d(k) is the minimum distance between the detected edge pixels and the actual edge points.

• Structural Similarity Index Measure (SSIM)

Structural Similarity Index Measure (SSIM) stands for Structural Similarity Index Measure. The equation is
given by:

(2µx µy + C1 )(2σxy + C2 )
SSIM(x, y) = (19)
(µ2x + µ2y + C1 )(σx2 + σy2 + C2 )

where µx and µy are the average intensities of the images (x) and (y), respectively. sigma2x and σy2 are the
variances of (x) and (y), respectively. σxy is the covariance of (x) and (y). C1 and C2 are small constants added
to avoid division by zero.

• Average Precision (AP)

The Average Precision metric (AP) is a performance evaluation metric used for object detection and informa-
tion retrieval models, which is given by:

TP
AP = (20)
TP + FP

Threshold function selection

The median threshold function is chosen for its stability and effectiveness. To verify the advantage of median
threshold, we use image “Butterfly” and wavelet“rbior 4.4”to extract the edge and calculate MSE, PSNR and
Accuracy using three threshold functions. The results is shown in Table 2.
The results obtained using the median thresholding method are comparable to those achieved with the other
two thresholding techniques. Under the Mean Squared Error (MSE) metric, the performance of the median
thresholding matches that of the hard thresholding, while it outperforms the soft thresholding. Similarly, in
terms of Peak Signal-to-Noise Ratio (PSNR), the median thresholding results are equivalent to those obtained
with hard thresholding and superior to those from soft thresholding. Furthermore, when evaluated using the

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 8

www.nature.com/scientificreports/

Soft Hard Median

MSE(pixel) 0.3619 0.3615 0.3615
PSNR(dB) 4.4136 4.4189 4.4189
Accuracy(%) 63.8060 63.8486 63.8496

Table 2. The MSE, PSNR and accuracy when different threshold function.

Fig. 6. Edge detection results of different denoising combinations at 0.1% Gaussian noise level: (a) Ground
truth image, (b) Canny, (c) bior4.4, (d) bl7, (e) coif5, (f) db5, (g) fk6, (h) haar, (i) han5.5, (j) mb4.2, (k) rbio4.4,
(l) sym5.

Accuracy metric, the median thresholding method surpasses both hard and soft thresholding techniques. Hence,
we selected median function in the experiment.

Edge extraction results

To assess the resilience of the adaptive edge extraction method, Gaussian noise levels of 0.1%, 10%, 20%, and
30% are introduced to the original image. The comparison is made between the Canny method and the Gaussian
and wavelet thresholding denoising method (GWM) to evaluate the efficacy of the proposed approach. The
low and high thresholds are determined using the modified OTSU method. Various denoising combination
methods (bior4.4, bl7, coif5, db5, fk6, haar, han5.5, mb4.2, rbio4.4, and sym5) are employed, and the resulting
edge extraction outcomes are presented. The test image is “Butterfly”.

• 0.1% Noise level

The adaptive edge extraction results at 0.1%noise level using various denoising methods are shown as Fig. 6.

• 10% Noise level

The adaptive edge extraction results at 10%noise level using various denoising methods are shown as Fig. 7.

• 20% Noise level

The adaptive edge extraction results at 20%noise level using various denoising methods are shown as Fig. 8.

• 30% Noise level

The adaptive edge extraction results at 30%noise level using various denoising methods are shown as Fig. 9.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 9

www.nature.com/scientificreports/

Fig. 7. Edge detection results of different denoising combinations at 10% Gaussian noise level: (a) Ground
truth image, (b) Canny, (c) bior4.4, (d) bl7, (e) coif5, (f) db5, (g) fk6, (h) haar, (i) han5.5, (j) mb4.2, (k) rbio4.4,
(l) sym5.

Fig. 8. Edge detection results of different denoising combinations at 20% Gaussian noise level: (a) Ground
truth image, (b) Canny, (c) bior4.4, (d) bl7, (e) coif5, (f) db5, (g) fk6, (h) haar, (i) han5.5, (j) mb4.2, (k) rbio4.4,
(l) sym5.

Figures 6, 7, 8, 9 illustrated that at a noise level of 0.1%, all wavelet denoising combinations achieved comparable
performance to the Canny method. Conversely, at noise levels of 10%, 20%, and 30%, the WTD scheme
exhibited superior edge detection effectiveness compared to the Canny method, extracting more accurate pixels
corresponding to the Ground Truth Image (GTI).

Evaluation
• MSE

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 10

www.nature.com/scientificreports/

Fig. 9. Edge detection results of different denoising combinations at 30% Gaussian noise level: (a) Ground
truth image, (b) Canny, (c) bior4.4, (d) bl7, (e) coif5, (f) db5, (g) fk6, (h) haar, (i) han5.5, (j) mb4.2, (k) rbio4.4,
(l) sym5.

0.1 10 20 30
Canny 0.3641 0.36552 0.37557 0.42185
bior4.4 0.3615 0.36119 0.35699 0.35802
bl7 0.36166 0.36049 0.35857 0.35643
coif5 0.36164 0.36192 0.36028 0.35988
db5 0.36166 0.36117 0.35921 0.35627
fk6 0.36166 0.35942 0.35338 0.35318
haar 0.36154 0.36196 0.36194 0.35839
han5.5 0.36162 0.36113 0.35701 0.356
mb4.2 0.36176 0.35966 0.34841 0.34762
rbio4.4 0.36141 0.35988 0.352 0.34968
sym5 0.36196 0.3636 0.36069 0.36069

Table 3. The MSE for different edge detection method at various noise levels.

The similarity between the extraction edge image and the GTI is measured by calculating the MSE. The results
of MSE are shown in Table 3 and Fig. 10a.
At low noise levels (0.1%), the performance of most wavelet functions is comparable to or slightly different
from that of the Canny wavelet. As the noise level increases, the performance differences among various
wavelet functions become more pronounced. Some wavelet functions, such as Bior4.4, Coif5, Han5.5, and
Sym5, exhibit better robustness, maintaining a lower mean squared error (MSE) even at high noise levels. In
contrast, the performance of the Canny wavelet deteriorates significantly at high noise levels, particularly at
a 30% noise level.

• PSNR

PSNR represents the ratio of the maximum possible power of a signal to the destructive noise power that
affects its representation accuracy. The results of PSNR are shown in Table 4 and Fig. 10b.
The GWM method demonstrates better robustness than the Canny method across different noise levels, es-
pecially in high-noise environments. At low noise levels, the performance difference between the GWM and
Canny methods is minimal; however, the performance improvement of the GWM method becomes more
significant at medium to high noise levels. In summary, the GWM method generally outperforms the Canny
method across various noise levels.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 11

www.nature.com/scientificreports/

Fig. 10. (a) The Mean Squared Error at different levels. (b) The Peak Signal to Noise Ration value at different
levels. (c) The Accuracy at different levels. (d) The Figure of Merit at different levels. (e) The Structural
Similarity Index Measure at different levels.

0.1 10 20 30
Canny 4.3878 4.41863 4.20707 3.74845
bior4.4 4.41887 4.42268 4.47347 4.46096
bl7 4.41697 4.43102 4.45424 4.48023
coif5 4.4172 4.41387 4.43365 4.43843
db5 4.41697 4.42292 4.44656 4.48216
fk6 4.41697 4.44393 4.51756 4.52
haar 4.41839 4.4134 4.41364 4.4564
han5.5 4.41744 4.42339 4.47323 4.48554
mb4.2 4.41578 4.44106 4.5791 4.58899
rbio4.4 4.42006 4.43843 4.53464 4.56333
sym5 4.4134 4.39371 4.42864 4.42864

Table 4. The PSNR for different edge detection method at various noise levels.

• Accuracy

The Accuracy determines the closeness of a measured or calculated value to its true or accepted value. The
results of Accuracy are shown in Table 5 and Fig. 10c.
At low noise levels (0.1%), the performance of the Canny method is comparable to that of the GWM method
using wavelet functions. As the noise level increases, specific wavelet functions (such as bior4.4, bl7, fk6,
han5.5, mb4.2, and rbio4.4) demonstrate a greater improvement in accuracy with the GWM method com-
pared to the Canny method, particularly under high noise conditions of 20% and 30%, where the perfor-
mance enhancement of these wavelet functions is more pronounced.

• Figure of Merit (FOM)

The Figure of Merit (FOM) is a quantitative measure used to evaluate the performance or efficiency. The
results of FOM is shown in Fig. 10d.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 12

www.nature.com/scientificreports/

0.1 10 20 30
Canny 63.5901 63.5076 62.8429 62.8153
bior4.4 63.8496 63.8813 64.3013 64.1983
bl7 63.8337 63.9506 64.1428 64.3567
coif5 63.8357 63.808 63.9724 64.012
db5 63.8337 63.8833 64.0794 64.3726
fk6 63.8337 64.0576 64.6618 64.6816
haar 63.8456 63.804 63.806 64.1606
han5.5 63.8377 63.8872 64.2993 64.4003
mb4.2 63.8238 64.0338 65.1591 65.2383
rbio4.4 63.8595 64.012 64.8005 65.0323
sym5 63.804 63.6396 63.9308 63.9308

Table 5. The accuracy for different edge detection method at various noise levels.

Under the processing of different wavelet functions, the FOM values generally increase as the noise level rises
from 0.1 to 30%, indicating an enhancement in denoising performance. Specifically, most wavelet functions
show an improvement in FOM values at the 30% noise level compared to the 0.1% noise level, suggesting that
these wavelet functions are more effective at preserving image quality when dealing with higher noise levels.
In particular, the Rbio4.4 wavelet function demonstrates the best denoising performance at higher noise levels
(20% and 30%), while the Canny wavelet shows relatively lower FOM values under all test conditions, indicat-
ing that its denoising performance is inferior to that of the other wavelet functions.

• Structural Similarity Index Measure (SSIM)

Structural Similarity Index Measure (SSIM) stands for Structural Similarity Index Measure. The results of
SSIM is shown in Fig. 10e.
The GWM method shows improvements over the Canny method for most wavelet functions. For instance,
when using the Bior4.4, Bl7, Coif5, Db5, Fk6, Haar, Han5.5, Mb4.2, Rbio4.4, and Sym5 wavelet functions,
the SSIM index is enhanced. At a noise level of 10%, the GWM method exhibits performance improvements
compared to the Canny method when using different wavelet functions. Even at the high noise level of 30%,
the performance enhancement of the GWM method relative to the Canny method can still be observed. Al-
though the performance of all methods may be affected under such extreme conditions, the GWM method
can still provide a certain degree of performance improvement when utilizing specific wavelet functions.
Overall, the GWM method demonstrates performance improvements over the Canny method across various
noise levels with the support of different wavelet functions. This enhancement indicates that the GWM meth-
od exhibits good robustness and effectiveness when handling different levels of noise.

• Running Time

To evaluate the efficiency and speed of the GWM methods, we assess the running time under various noise
levels. The curve is shown in Fig. 11 and the table of average time is shown in Table 6.
At a low noise level of 0.1%, the haar, han5.5, and rbio4.4 wavelet functions not only demonstrated high
performance, but according to the data in the table, the average processing times were 475.48 ms for Haar,
504.60 ms for han5.5, and 500.09 ms for rbio4.4. These wavelet functions maintained relatively short average
processing times while achieving high performance, indicating that they are effective and efficient in low
noise environments.
As the noise level increases, we observed that the bior4.4 and fk6 wavelet functions performed relatively well
at high noise levels of 30%, with average processing times of 479.21 ms and 430.15 ms, respectively. This
indicates that despite the increased difficulty of denoising in high noise environments, these two wavelet
functions can still maintain relatively good performance and reasonable processing times.

• Discussions

The analysis reveals that the edge detection outcomes of the GWM outperform those of the Canny method
across various levels of Gaussian noise (0.1%, 10%, 20%, and 30%).
The GWM variations (bior4.4, bl7, coif5, db5, fk6, haar, han5.5, mb4.2, rbio4.4, and sym5) exhibit superior
MSE, PSNR, and Accuracy compared to the Canny method exclusively. These results demonstrate the robust-
ness and high precision of EDAW approach to Gaussian noise at different levels.
The EDAW method demonstrates better performance than the Canny method across five metrics-MSE, Ac-
curacy, PSNR, FOM, and SSIM-when handling images at different noise levels, particularly in high-noise
environments. This indicates that the GWM method has a clear advantage in image denoising and feature
preservation. However, the Canny method performs more consistently in low-noise environments, but its

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 13

www.nature.com/scientificreports/

Fig. 11. The running time of GWM methods.

bior4.4 bl7 coif5 db5 fk6 haar han5.5 mb4.2 rbio4.4 sym5
Time (ms) 479.21 363.72 539.11 391.76 430.15 475.48 504.60 363.81 500.09 365.63

Table 6. The average running time at various noise levels.

performance deteriorates under high noise conditions, especially in preserving image structure and shape
features. Therefore, the EDAW method is more effective for processing high-noise images, while the Canny
method may be more suitable for low-noise environments.

The EDAW method demonstrated superior performance compared to using the Canny method. This may be
attributed to three main factors:

1. Gaussian denoising with excessively large sigma values can overly smooth image de-tails, while using too
small sigma values may result in inadequate smoothing. Combining a Gaussian filter with a small sigma and
WTD can achieve optimal performance.
2. Gaussian filtering uniformly smooths the entire image, potentially blurring important along with. contrast,
the WTD method decomposes the image into different frequency bands, analyzes the coefficients, and effec-
tively distinguishes noise from actual image features. This strategy allows for the suppression or removal of
noise coefficients based on a thresholding approach, resulting in superior denoising outcomes compared to
Gaussian filtering.
3. OTSU’s adaptive thresholding can effectively separate objects of interest from the background, leading to
improved precision and accuracy in image processing tasks.

Comparison results
In this comparative experiment, we assess the performance of our proposed EDAW against several established
methods, including traditional techniques like Canny, Roberts, Sobel, and SLEDGE57, as well as deep learning
approaches such as RCF58 and the method by Li et al.59 The experiments utilize the BSDS500 benchmark
dataset56, known for its diverse images and ground truth annotations. EDAW employs the fk6 wavelet function,
which preliminary tests show yields the highest PSNR at a noise level of 0.1%. The Canny and Roberts functions
are derived from the matlab function library.
We present visual comparisons of edge detection results (Fig. 12) and Average Precision (AP) scores (7) for
each algorithm to assess their performance quantitatively.
It can be seen from Fig. 12, The EDAW method achieves more complete edges compared to the Canny and
Roberts algorithms. However, the edges extracted by all three algorithms still exhibit certain differences from the
ground truth, which is due to the limitations of traditional methods. Notably, these traditional edge detection
algorithms do not incorporate labeled training data, which can restrict their performance in complex scenarios.
As seen in Table 7, our proposed method, EDAW, achieved an impressive Average Precision (AP) score of
0.682, significantly outperforming traditional methods. For instance, the Canny edge detector attained an AP
of only 0.58, while the Sobel operator scored 0.568, and the Roberts method lagged behind with an AP of 0.512.
The SLEDGE method followed closely, securing second place among traditional approaches with an AP of 0.652.
In the realm of deep learning techniques, the RCF method led with a notable AP of 0.822, and the method
developed by Li et al. achieved an AP of 0.605, placing it second among deep learning approaches. These
results underscore that, while deep learning methods generally excel, our EDAW approach distinctly surpasses

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 14

www.nature.com/scientificreports/

Fig. 12. The visual comparisons of edge detection results.

Methods AP
Canny 0.580
Sobel 0.568
Roberts 0.512
SLEDGE 0.652
RCF 0.822
Li et al. 0.605
EDAW 0.682

Table 7. The average precision (AP) scores for each algorithm.

traditional techniques and approaches the performance levels of deep learning solutions. This demonstrates the
significant competitiveness and potential of our proposed method in edge detection tasks.
The superior performance of EDAW can be attributed to several key innovations. Firstly, the integration
of a denoising module that employs wavelet and Gaussian denoising techniques enhances image quality by
effectively reducing noise interference. This improves edge detection accuracy, resulting in a higher AP score.
Moreover, the implementation of an adaptive thresholding method based on a modified OTSU algorithm
allows for flexible, context-sensitive edge detection. This approach dynamically adjusts the thresholding
criteria based on varying image characteristics, further enhancing the ability to differentiate between edges and
background noise. The incorporation of median thresholding, known for its stability, strengthens the robustness
of our method and mitigates potential artifacts associated with hard thresholding techniques.
In contrast, traditional methods such as Canny and Roberts exhibit inherent limitations. The Canny edge
detector, while popular, is sensitive to noise and heavily reliant on parameter selection, which can lead to
suboptimal performance in images with variable noise levels, as evidenced by its AP of 0.58 in our experiments.
The Roberts operator, designed to detect edges through local intensity changes, suffers from its simplistic
approach, achieving a low AP of 0.512 and struggling with complex edge structures and noise.
Additionally, the detector proposed by Li et al. faces challenges due to its reliance on training data that may
not align with the specific requirements of contour detection. Its design is influenced by local color changes,
which fails to capture the broader contextual information necessary for accurate contour detection.
Overall, the systematic design of EDAW, particularly its focus on adaptive thresholding and robust denoising,
positions it as a superior method in the field of edge detection. This is particularly evident when compared
to traditional methods like Canny and Roberts, as well as certain neural network-based approaches that may
not be optimally trained or designed for edge detection tasks. These findings highlight the importance of
tailored methodologies in addressing the unique challenges of edge and contour detection across diverse image
conditions.

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 15

www.nature.com/scientificreports/

Conclusion
The EDAW methodology we proposed, which combines adaptive thresholding and wavelet denoising, provides
a robust and automated approach to edge detection. By integrating wavelet and Gaussian denoising with
automated selection techniques, EDAW effectively eliminates noise from images and accurately identifies edge
pixels. This method demonstrates high precision detection capabilities through adaptive threshold selection
and edge connection, as evidenced by performance comparisons with recent methods across various levels of
Gaussian noise.
To assess the edge extraction implementation, six metrics (MSE, Accuracy, PSNR, FOM, SSIM and Running
time) are utilized. Experimental results indicate that the performance of EDAW surpasses that of using the
Canny method at different Gaussian noise levels (0.1%, 10%, 20%, and 30%). Comparative results also indicate
that EDAW outperforms traditional methods, such as Canny and Roberts showcasing its effectiveness in edge
detection.
Despite its strong performance, EDAW may still face challenges in extremely noisy environments or with
complex image content. Future research should prioritize developing adaptive mechanisms for parameter
selection to bolster its robustness across diverse scenarios. By addressing these limitations and exploring new
practical applications, EDAW has the potential to advance the field of edge detection and enhance image analysis
in various domains.
Looking ahead, future work will focus on enhancing the computational efficiency of EDAW and its
adaptability to various image types, particularly for high-resolution images essential for real-time applications.
Integrating EDAW into smartphone and digital camera systems could significantly improve user experience
in consumer photography and videography by enabling high-quality edge detection, allowing users to capture
sharper images in challenging conditions.

Data availability
The datasets used and analysed during the current study available from the corresponding author on reasonable
request.

Received: 17 August 2024; Accepted: 14 January 2025

References
1. Zhou, G. et al. Orthorectification of fisheye image under equidistant projection model. Remote ​S​e​n​s​.​[​S​P​A​C​E​]​​h​t​t​p​s​:​/​/d
​ ​o​i​.​o​r​g​/​1​0.​ ​3​3​
9​0​/​r​s​1​41​ ​7​4​1​7​5​​​​ (2022).
2. Ghandorh, H. et al. Semantic segmentation and edge detection-approach to road detection in very high resolution satellite images.
Remote Sens.[SPACE]https://doi.org/10.3390/rs14030613 (2022).
3. Wang, J., Chen, G., Zhang, T. J., Wu, N. & Wang, X. An efficient muscle segmentation method via Bayesian fusion of probabilistic
shape modeling and deep edge detection. IEEE Trans. Biomed. Eng.[SPACE]https://doi.org/10.1109/TBME.2024.3415818 (2024).
4. Zhou, G., Li, H. & Gao, E. Overview on interest point detection over 40 year’s development: A review. IEEE Sens. J. 23, 26942–
26951. https://doi.org/10.1109/JSEN.2023.3318756 (2023).
5. Wang, M., Sun, C. & Sowmya, A. Efficient corner detection based on corner enhancement filters. Digit. Signal Process. 122, 103364.
https://doi.org/10.1016/j.dsp.2021.103364 (2022).
6. Wu, F., Zhu, C., Xu, J., Bhatt, M. W. & Sharma, A. Research on image text recognition based on canny edge detection algorithm and
k-means algorithm. Int. J. Syst. Assur. Eng. Manag. 13, 72–80 (2022).
7. Lu, Y., Duanmu, L., Zhai, Z. J. & Wang, Z. Application and improvement of canny edge-detection algorithm for exterior wall
hollowing detection using infrared thermal images. Energy Build. 274, 112421. https://doi.org/10.1016/j.enbuild.2022.112421
(2022).
8. Ji, G.-P., Zhu, L., Zhuge, M. & Fu, K. Fast camouflaged object detection via edge-based reversible re-calibration network. Pattern
Recognit. 123, 108414. https://doi.org/10.1016/j.patcog.2021.108414 (2022).
9. Rani, S., Ghai, D. & Kumar, S. Object detection and recognition using contour based edge detection and fast r-CNN. Multimedia
Tools Appl. 81, 42183–42207 (2022).
10. Huang, Z. et al. Making accurate object detection at the edge: Review and new approach. Artif. Intell. Rev. 55, 2245–2274 (2022).
11. He, C. et al. Camouflaged object detection with feature decomposition and edge reconstruction. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 22046–22055 (2023).
12. Yang, Z., Soltanian-Zadeh, S. & Farsiu, S. Biconnet: An edge-preserved connectivity-based approach for salient object detection.
Pattern Recognit. 121, 108231. https://doi.org/10.1016/j.patcog.2021.108231 (2022).
13. Wu, Y. et al. Edge computing driven low-light image dynamic enhancement for object detection. IEEE Trans. Netw. Sci. Eng. 10,
3086–3098. https://doi.org/10.1109/TNSE.2022.3151502 (2023).
14. Pavithra, L. & Sharmila, T. S. An efficient framework for image retrieval using color, texture and edge features. Comput. Electr. Eng.
70, 580–593. https://doi.org/10.1016/j.compeleceng.2017.08.030 (2018).
15. Jankowski, M., Gündüz, D. & Mikolajczyk, K. Wireless image retrieval at the edge. IEEE J. Select. Areas Commun. 39, 89–100.
https://doi.org/10.1109/JSAC.2020.3036955 (2021).
16. Jing, J., Liu, S., Wang, G., Zhang, W. & Sun, C. Recent advances on image edge detection: A comprehensive review. Neurocomputing
503, 259–271. https://doi.org/10.1016/j.neucom.2022.06.083 (2022).
17. Sun, R. et al. Survey of image edge detection. Front. Signal Process. 2, 1. https://doi.org/10.3389/frsip.2022.826967 (2022).
18. Lyu, C. et al. Seeing the vibration: Visual-based detection of low frequency vibration environment pollution. IEEE Sens. J. 21,
10073–10081. https://doi.org/10.1109/JSEN.2021.3059110 (2021).
19. Krishnan Nair, S. et al. Prewitt logistic deep recurrent neural learning for face log detection by extracting features from images.
Arab. J. Sci. Eng. 48, 2589 (2023).
20. Canny, J. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell. PAMI–8, 679–698. ​h​t​t​p​s​:​/​/​d​o​i​.​or​ ​g​/​
1​0​.​1​1​0​9/​ ​T​P​A​M​I​.​1​9​86​ ​.​4​7​6​7​8​5​1​​​​ (1986).
21. Xuan, L. & Hong, Z. An improved canny edge detection algorithm. In 2017 8th IEEE International Conference on Software
Engineering and Service Science (ICSESS), 275–278 (2017). https://doi.org/10.1109/ICSESS.2017.8342913.
22. Patil, R. V. & Reddy, Y. An autonomous technique for multi class weld imperfections detection and classification by support vector
machine. J. Nondestruct. Eval. 40, 76 (2021).

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 16

www.nature.com/scientificreports/

23. Khmag, A., Al Haddad, S. A. R., Ramlee, R. A., Kamarudin, N. & Malallah, F. L. Natural image noise removal using nonlocal means
and hidden Markov models in transform domain. Vis. Comput. 34, 1661–1675 (2018).
24. Ofir, N. et al. On detection of faint edges in noisy images. IEEE Trans. Pattern Anal. Mach. Intell. 42, 894–908. ​h​t​t​ps​ ​:​/​/​d​o​i​.​o​rg​ ​/​1​0​.​1​
1​0​9​/T ​ ​P​A​M​I​.​2​0​1​9.​ ​2​8​9​2​1​3​4​​​​ (2020).
25. Marr, D. & Hildreth, E. Theory of edge detection. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 207, 187–217 (1980).
26. Yang, W., Chen, X.-D., Wang, H. & Mao, X. Edge detection using multi-scale closest neighbor operator and grid partition. Vis.
Comput. 40, 1947–1964 (2024).
27. Khmag, A. Additive gaussian noise removal based on generative adversarial network model and semi-soft thresholding approach.
Multimed. Tools Appl. 82, 7757–7777 (2023).
28. Khmag, A. Natural digital image mixed noise removal using regularization Perona–Malik model and pulse coupled neural
networks. Soft Comput. 27, 15523–15532 (2023).
29. Bradley, D. & Roth, G. Adaptive thresholding using the integral image. J. Graph. Tools 12, 13–21. ​h​t​t​p​s​:​/​/​d​o​i​.​or​ ​g​/​1​0​.​1​0​8​0/​ ​2​1​5​1​2​3​7​
X​.2​ ​0​0​7​.​1​0​1​2​9​2​3​6​​​​ (2007).
30. Wang, G., Lopez-Molina, C. & De Baets, B. Multiscale edge detection using first-order derivative of anisotropic gaussian kernels.
J. Math. Imaging Vis. 61, 1096–1111 (2019).
31. Mafi, M., Rajaei, H., Cabrerizo, M. & Adjouadi, M. A robust edge detection approach in the presence of high impulse noise
intensity through switching adaptive median and fixed weighted mean filtering. IEEE Trans. Image Process. 27, 5475–5490. ​h​t​t​p​s​:​/​
/​d​o​i​.​o​r​g​/​1​0.​ ​1​1​0​9​/​T​I​P​.2​ ​0​1​8​.​2​8​5​7​44​ ​8​​​​ (2018).
32. Cao, J., Chen, L., Wang, M. & Tian, Y. Implementing a parallel image edge detection algorithm based on the OTSU-canny operator
on the Hadoop platform. Comput. Intell. Neurosci. 2018, 3598284 (2018).
33. You, N. et al. Research on wavelet transform modulus maxima and OTSU in edge detection. Appl. ​S​c​i​.​[​S​P​A​C​E​]​​h​t​t​p​s​:​/​/​d​o​i​.​o​rg​ ​/​1​0​.​
3​3​9​0​/a​ ​p​p​1​3​0​7​4​4​5​4​​​​ (2023).
34. Akagic, A., Buza, E., Omanovic, S. & Karabegovic, A. Pavement crack detection using OTSU thresholding for image segmentation.
In 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO),
1092–1097 (2018). https://doi.org/10.23919/MIPRO.2018.8400199.
35. Feng, Y. et al. A novel interval iterative multi-thresholding algorithm based on hybrid spatial filter and region growing for medical
brain MR images. Appl. Sci.[SPACE]https://doi.org/10.3390/app13021087 (2023).
36. Guo, L. & Wu, S. FPGA implementation of a real-time edge detection system based on an improved canny algorithm. Appl.
Sci.[SPACE]https://doi.org/10.3390/app13020870 (2023).
37. Hong, S. et al. Improved mask r-CNN combined with OTSU preprocessing for rice panicle detection and segmentation. Appl.
Sci.[SPACE]https://doi.org/10.3390/app122211701 (2022).
38. Jin, J., Zhou, W., Yang, R., Ye, L. & Yu, L. Edge detection guide network for semantic segmentation of remote-sensing images. IEEE
Geosci. Remote Sens. Lett. 20, 1–5. https://doi.org/10.1109/LGRS.2023.3234257 (2023).
39. Soria, X., Sappa, A., Humanante, P. & Akbarinia, A. Dense extreme inception network for edge detection. Pattern Recognition 139,
109461. https://doi.org/10.1016/j.patcog.2023.109461 (2023).
40. Soria, X., Sappa, A., Humanante, P. & Akbarinia, A. Dense extreme inception network for edge detection. Pattern Recognition 139,
109461. https://doi.org/10.1016/j.patcog.2023.109461 (2023).
41. Khmag, A. & Kamarudin, N. Natural image deblurring using recursive deep convolutional neural network (r-dbcnn) and second-
generation wavelets. In 2019 IEEE International Conference on Signal and Image Processing Applications (ICSIPA), 285–290 (2019).
https://doi.org/10.1109/ICSIPA45851.2019.8977756.
42. Gonzales, R. C. & Wintz, P. Digital Image Processing 2nd edn. (Addison-Wesley Longman Publishing Co., Inc., USA, 1987).
43. Minaee, S. et al. Image segmentation using deep learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 44, 3523–3542. ​ht​ ​t​p​s​:​/​
/​d​o​i.​ ​o​r​g​/​1​0​.​1​10​ ​9​/​T​P​A​M​I​.​20​ ​2​1​.​3​0​5​9​9​6​8​​​​ (2022).
44. Ni, B., Song, F., Zhao, L., Fu, Z. & Huang, Y. Wavelet denoising of fiber optic monitoring signals in permafrost regions. Sci. Rep. 14,
9085 (2024).
45. Tian, C. et al. Multi-stage image denoising with the wavelet transform. Pattern Recognit. 134, 109050. ​h​t​t​p​s​:/​ ​/​d​o​i​.​o​r​g​/1​ ​0​.​1​0​1​6​/​j​.p ​ ​a​
t​c​o​g​.​2​0​22​ ​.​1​0​9​0​5​0​​​​ (2023).
46. Cohen, A., Daubechies, I. & Feauveau, J.-C. Biorthogonal bases of compactly supported wavelets. Commun. Pure Appl. Math. 45,
485–560. https://doi.org/10.1002/cpa.3160450502 (1992).
47. Antoniadis, A. & Oppenheim, G. Wavelets and Statistics Vol. 103 (Springer, Berlin, 2012).
48. Donoho, D. L. Progress in wavelet analysis and WVD: A ten minute tour. In Progress in Wavelet Analysis and Applications 109–128
(1993).
49. Donoho, D. L. & Johnstone, I. M. Ideal spatial adaptation by wavelet shrinkage. Biometrika 81, 425–455. ​h​t​t​p​s​:​/​/d ​ ​o​i​.​o​r​g​/​1​0.​ ​1​0​9​3​/​b​
i​o​me​ ​t​/​8​1​.​3​.​4​2​5​​​​ (1994).
50. Donoho, D. De-noising by soft-thresholding. IEEE Trans. Inf. Theory 41, 613–627. https://doi.org/10.1109/18.382009 (1995).
51. Donoho, D. L., Johnstone, I. M., Kerkyacharian, G. & Picard, D. Wavelet shrinkage: Asymptopia?. J. R. Stat. Soc. Ser. B (Methodol.)
57, 301–337. https://doi.org/10.1111/j.2517-6161.1995.tb02032.x (2018).
52. Johnstone, I. M. & Silverman, B. W. Needles and straw in haystacks: Empirical Bayes estimates of possibly sparse sequences. Ann.
Stat. 32, 1594–1649. https://doi.org/10.1214/009053604000000030 (2004).
53. Abramovich, F., Benjamini, Y., Donoho, D. L. & Johnstone, I. M. Adapting to unknown sparsity by controlling the false discovery
rate. Ann. Stat. 34, 584–653. https://doi.org/10.1214/009053606000000074 (2006).
54. You, N., Han, L., Zhu, D. & Song, W. Research on image denoising in edge detection based on wavelet transform. Appl.
Sci.[SPACE]https://doi.org/10.3390/app13031837 (2023).
55. Piosenka, G. Butterfly and moths image classification 100 species (2022). ​h​t​t​p​s​:​​​/​​/​w​w​​w.​ ​k​a​g​g​l​​e​.​co
​​ ​m​​/​d​a​t​a​s​​e​t​s​/​g​​p​i​o​s​e​n​​k​a​/​b​u​​t​t​e​r​​f​l​y​-​i​
m​a​g​e​s​4​0​-​s​p​e​​ci​ ​e​s​/​d​a​t​a.
56. Arbeláez, P., Maire, M., Fowlkes, C. & Malik, J. Contour detection and hierarchical image segmentation. IEEE Trans. Pattern Anal.
Mach. Intell. 33, 898–916. https://doi.org/10.1109/TPAMI.2010.161 (2011).
57. Payet, N. & Todorovic, S. Sledge: Sequential labeling of image edges for boundary detection. Int. J. Comput. Vis. 104, 15–37 (2013).
58. Liu, Y., Cheng, M.-M., Hu, X., Wang, K. & Bai, X. Richer convolutional features for edge detection. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (CVPR) (2017).
59. Li, O. & Shui, P.-L. Color edge detection by learning classification network with anisotropic directional derivative matrices. Pattern
Recognit. 118, 108004. https://doi.org/10.1016/j.patcog.2021.108004 (2021).

Acknowledgements
This paper is financially supported by the National Natural Science of China (the Grant: 42474039) and the
Innovative research team of Henan Polytechnic University (the grant:T2024-3).

Author contributions
Conceptualization, H.L.; methodology, H.L. and K.X.; software, H.L.; validation, H.L. and K.X.; formal analysis,
H.L.; investigation, H.L.; resources, K.X.; data curation, H.L.; writing-original draft preparation, H.L.; writing

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 17

www.nature.com/scientificreports/

review and editing, H.L. and K.X.; visualization, H.L.; supervision, K.X.; project administration, K.X.; funding
acquisition, K.X. All authors have read and agreed to the published version of the manuscript.

Declarations

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to K.X.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives
4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in
any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide
a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have
permission under this licence to share adapted material derived from this article or parts of it. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence
and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to
obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​p​:​/​/​c​re​ ​a​t​i​v​e​c​o​m​mo
​ ​
n​s​.​o​r​g​/​l​ic​ ​e​n​s​e​s​/​b​y​-n
​ ​c​-​n​d​/​4​.​0​/​​​​.​​

© The Author(s) 2025

Scientific Reports | (2025) 15:5838 | https://doi.org/10.1038/s41598-025-86860-9 18