Sobel-Canny Fusion for Effective Edge Detection in Gaussian

Denoised Digital Images

Sahana R

1,2

, Manjula Gururaj Rao

and Ganesh Aithal

Dept. of CSE, NTTTE (Deemed to be University), NMAMIT, Nitte, Karkala, India

Dept. of CSE, SMVITM, Bantakal, India

NTTTE (Deemed to be University), NMAMIT, Nitte, Karkala, India

Keywords: Gaussian, PSNR, Time, Sobel and Canny.

Abstract: Information is essential in the digital age. These data may take the shape of pictures or statistics. Getting

high-quality photos is crucial when working with image data. Unwanted information in an image is

referred to as noise, and it presents a major difficulty for image analysis. Depending on the image format,

many types of noise may exist. Image noise removal is a difficult process. The suggested effort focusses

on removing noise from colored images, especially from video frames. To improve the quality of the

frames by lowering noise, the model uses Non-Local Means Denoising, Gaussian Filter, and Bilateral

Filter approaches. Time performance and PSNR (peak signal-to-noise ratio) measurements are used to

assess how effective the techniques are. The final photos had substantially higher PSNR values. One

important component of digital image/video processing is edge detection. Edge detection is essential for

applications that required to extract features or object information from a picture. Although there are

several different edge detection operators are available today, improving the performance of current

system remains a difficulty In this study, the Sobel and Canny edge detectors are combined in a hybrid

technique.

1 INTRODUCTION

Digital video is used in many applications like

medical, entertainment, information technology,

multimedia services etc. Both TV shows and movies

have switched to digital and high definition,

moreover now the majority of movies and certain

TV shows are in 3D format. Digital video unites the

movies, communication sector and computer,

making phone, cable TV and internet providers

bitter rivals. One device can function as phone, HD

TV and personal computer. These days we can use

mobile to record live video, a laptop to handle

digital data and printer to produce still image.

Enhancing the image quality is an ongoing area

of research. As low-end gadgets like mobile phones

and webcams proliferate, digital image and video

enhancements methods are increasingly needed to

improve their outputs. The main cause of image

quality degradation is noise. Thus, it is essential to

explore denoising algorithm to generate high quality

images and video frames. (Z. Liu,2014), ( M.

Maggioni,2012), (Mildenhall,2018) This paper

focuses on video denoising. Here the objective is to

develop high quality video denoising algorithm, that

successfully eliminate noise introduced by digital

cameras.

Low light, sensor flaws and thermal impacts are

the causes of noise in the images. It requires lot of

time and effort to manually adjust several

parameters in order to enhance the quality and

bandwidth of cameras. In real world scenario, where

color and light conditions may vary abruptly

creating a distinct noise distribution for every frame,

it is required to have noise reduction algorithm that

operates on blind settings. The majority of

conventional denoising techniques choose and

average the images pixels to provide good results.

The fundamental components of various denoising

techniques are how to efficiently choose appropriate

pixels and calculate the averaging weights. Filters

can be used to improve and reduce the noise in

images and video frames. (Antoni Buades, 2011) A

linear filter type based on gaussian function is called

gaussian filter. The bilateral filter is however non-

R, S., Gururaj Rao, M. and Aithal, G.

Sobel-Canny Fusion for Effective Edge Detection in Gaussian Denoised Digital Images.

DOI: 10.5220/0013608400004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 41-48

ISBN: 978-989-758-763-4

linear filter type. It eliminates the noise while

maintaining the edge. Edge detection is essential for

applications that required to extract features or

object information from a picture. Although there

are several different edge detection operators are

available today, improving the performance of

current system remains a difficulty. In this study, the

Sobel and Canny edge detectors are combined in a

hybrid technique.

The section 2 focuses on the work done in the

removal of the noise from the images and proposed

work, implementation and the results are discussed

in the sections 3, 4 and 5 respectively

2 RELATED WORK

Denoising and edge detection of images and videos

has been researched for many years. Since

providing a comprehensive analysis is beyond the

scope of this paper, we will concentrate on

examining the work that is most relevant to ours.

Recently concept of image sparsity has been

introduced as image self-similarity, meaning that

patches within image exhibit similarities to each

other, which give rise to non-local methods

(Buades,2005). In NLM similar patches combined

with the weights that reflect with their similarities.

This straight forward approach yields high quality

outcomes. NLM techniques tests the images with

different noise levels. According to experimental

data this algorithm enhances the denoising

performances (Lingli Huang,2015). For this reason,

we have selected NLM as the one of the methods in

our denoising system.

One of the linear filters called gaussian filter

used to smooth and remove the noise from

images(M. G. Rao H,2019). Test and evaluation

report indicates that, it can be applied to enhance the

image quality (Sriani,2022) (Kumar,2020). The

gaussian filter approach is particularly useful for

filtering images with lot of noise, since filtering

finding shown a robust dependence on gaussian

kernel and relative independence on noise features

(Priyanka, 2020).

A nonlinear technique for denoising images

while maintaining the sharp edges is bilateral

filtering. The weighted sum of each pixels neighbors

in the input image determines its value in the output

image (P. D. Patil, 2015). It performs better than

linear filter like mean filter, weiner filter. In high

frequency area it performs better at eliminating the

noise, while in low frequency area it is ineffective

(Bhonsle,2012).

One important component of digital image/video

processing is edge detection. The performance of

each edge detection techniques is examined through

comparison. The findings demonstrated that in

contrast to Roberts and Prewitt operators, Sobel and

Canny edge detectors are less sensitive to random

noise in an image (Amer,2015) It is also suggested

to combine the canny and sobel operator for edge

detection (A. Kalra,2016).

In the proposed work focuses on the removal of

the noise using non local means denoising, Gaussian

filter and bilateral filtering. These algorithms are

compared and assessed using two kinds of metrics.

The PSNR and time performance are used as

evaluation metrics to determine the optimal filter to

reduce noise from frame under various conditions.

The Gaussian filter has been proven to be

superior in both the situations, however it does not

retain the edges for subsequent processing. To get

around this, hybrid approaches have been developed

to identify the edges in images while using the

Gaussian noise removal methods.

3 PROPOSED WORK

Three different methods have been implemented to

carry out denoising.

1. Non-Local Means Denoising

2. Gaussian Filter

3. Bilateral Filter

These methods perform denoising on each frame

and writes the denoised frames to an output video

3.1 Non-Local Means Denoising

The fundamental principle of Non local means

denoising is to substitute the average color of

neighboring pixels for pixel’s original color. In

probability theory, the variance law guarantees that

the noise standard deviation of an average of nine

pixels is divided by three. Therefore, one can split

the noise by three (and four with sixteen identical

pixels so on), if we can locate nine other pixels in

the image that are the same color as each pixel. To

identify all the pixels that actually resemble the pixel

needing denoising, it is acceptable to scan a large

area of the image. After that, denoising is

accomplished by calculating the average color of

these pixels that are most similar. Instead of

focusing only on color, the likeliness is assessed by

INCOFT 2025 - International Conference on Futuristic Technology

Figure 1: Proposed Algorithm.

analyzing the entire window surrounding each pixel.

We refer to this filter as non-local means

(Buades,2006) (Buades,2005) and it is given by

equation (1).

𝑁𝐿𝑢

(

𝑝

)

𝐶

(

𝑝

)

𝑓(𝑑𝐵

(

𝑝

)

,𝐵

(

𝑞

)

𝑢

(

𝑞

)

𝑑𝑞 (1)

The Euclidean distance between image centered

between p and q respectively is represented as

d(B(p), B(q)), f is a decreasing function, C(p) is the

normalizing factor.

3.2 Gaussian Filter

The most common form of noise in digital image

or video is Gaussian noise, while there are other

varieties as well such as impulse noise, speckle noise

and pepper noise. The gaussian noise in image/video

arises from sensor constraints’ during low light

conditions, which hinder the light sensor’s ability to

effectively record scene details. The bivariate

circular Gaussian function equation (2) can be used

mathematically to describe gaussian noise as follows

𝑔

(

𝑥,𝑦

)











𝑒





(





)























(



)

Where 𝜎

,

𝜎



are standard deviations and 𝜇



,𝜇



are the means.

In general, the output of the gaussian filter,

which uses the weighting function with a gaussian

distribution, is an average of the pixel values within

a specific neighborhood. This method involves

reducing the lot of high frequency noise such as

edges and small features to produce a smoother

image. The degree of smoothing function is

determined by Gaussian function standard deviation

or sigma

• Small sigma: They preserve more of the

surfaces and objects characteristics and

details since they have less smoothing

• Large sigma: A bit more smoothing, even

capable of erasing the key characteristics

Figure 2 depicts the relationship between

function value and standard deviation. Additionally

the function’s maximum value decreases as standard

deviation increases, other values deviating from

mean also increases (Gonzalez, 2007).

Figure 2: The graph shows variation in function value

according to the sigma value

3.3 Bilateral Filter

Bilateral filter is defined as weighted average of

adjacent pixels, much like gaussian filter. The

bilateral filter differs in that it preserves the edges

during smoothing by considering the difference in

value with neighbors. The fundamental principle of

this filter is that a pixel must have a similar value in

addition to occupying a nearby location, in order to

affect another pixel. The bilateral filter denoted by

BF[] is given by equation (3)

𝐵𝐹



𝐼











∑

𝐺









𝑝−𝑞

𝐺





𝐼



−𝐼



𝐼



(3)

Sobel-Canny Fusion for Effective Edge Detection in Gaussian Denoised Digital Images

Where normalization factor 𝑊



guarantees pixel

weights sum to 1.0

𝑊



=𝐺









𝑝−𝑞

𝐺





𝐼



−𝐼





(

)

The amount of filtering applied to the picture is

determined by 𝜎



𝑎𝑛𝑑 𝜎



. 𝐺





is a spatial gaussian

weighting and 𝐺





is a range gaussian.

Two parameters 𝜎



𝑎𝑛𝑑 𝜎



controls the bilateral

filter

• As the parameter 𝜎



increases, the bilateral

filter progressively approaches gaussian

filter more closely because 𝐺





flattens and

widens

• Larger features are smoothed by increasing

parameter 𝜎



For comparing between the above three algorithms,

we used two evaluation metrics.

i. The first metric is time. The time taken by

the whole program is calculated

ii. The second metric is PSNR. A standard

metric for assessing the quality of any video

or image is PSNR. In general, higher PSNR

denotes the higher quality reconstruction.

The formula given in below equation (5) is

used to determine the videos PSNR

𝑃𝑆𝑁𝑅 = 20log



𝑀𝐴𝑋



−10log



𝑀𝑆𝐸

(

)

MSE is mean squared error which is

calculated using below formula

𝑀𝑆𝐸 =

𝑚𝑛





𝐼

(

𝑖,𝑗

)

−𝐾

(

𝑖,𝑗

)





(

)









Where I is the original Image, K is the

denoised image, 𝑀𝐴𝑋



is the maximum

possible pixel value

It has been discovered that in both case Gaussian

filter outperform compared to other two algorithms.

Gaussian filter on the other hand do not maintain the

edge when eliminating noise. Gaussian denoising is

not ideal for further image or video processing as it

does not effectively preserve the edges. Hybrid

approaches have been put forth to improve edge

identification in gaussian denoised frame in order to

overcome this constraint. Specifically, combination

of Sobel and canny edge detection algorithms is

employed to accurately identify the edges while

mitigating the effects of gaussian noise.

One important component of digital image/video

processing is edge detection. Edge detection is

essential for applications that required to extract

features or object information from a picture. Edge

happens when the intensity functions abruptly

change or becomes discontinuous. Although there

are several different edge detection operators are

available today, improving the performance of

current system remains a difficulty. In this study, the

Sobel and Canny edge detector are combined in a

hybrid technique. Additionally, gaussian filter

eliminates noise from images and video when edge

identification is challenging by using it.

3.4 Sobel Edge Detector

The first order differential is the foundation of the

Sobel operator. This operation applies small,

separable integer valued filter to the image in both x

and y directions, making it computationally

inexpensive. It functions as an orthogonal gradient

operator, calculating the partial derivatives in both x

and y directions using 3x3 neighborhood of f(x,y)

(Moslem, 2011). The gradient in oblique direction is

given by equation (8)

𝑔

(

𝑥,𝑦

)

𝑆



+𝑆





(

)

𝑤ℎ𝑒𝑟𝑒 𝑆



{

𝑓

(

𝑥+1,𝑦−1

)

+2𝑓

(

𝑥+1,𝑦

)

𝑓

(

𝑥+1,𝑦+1

)



−{𝑓

(

𝑥−1,𝑦−1

)

2𝑓

(

𝑥−1,𝑦

)

+𝑓

(

𝑥+1,𝑦−1

)

(8)

𝑎𝑛𝑑 𝑆



{

𝑓

(

𝑥−1,𝑦+1

)

+2𝑓

(

𝑥,𝑦+1

)

𝑓

(

𝑥+1,𝑦+1

)



−{𝑓

(

𝑥−1,𝑦−1

)

2𝑓

(

𝑥,𝑦−1

)

+𝑓(𝑥−1,𝑦+1) (9)

Sobel edge detection has 2 benefits.

i) It smooths out some of the undesired

noise in a picture.

ii) By using differential of two rows and two

columns, edge’s element are strengthened,

making it thicker and brighter

3.5 Canny Edge Detector

A multi stage technique for identifying the different

edges in images or videos is called as canny edge

detector. Due its excellence performance, it is

sometimes referred as ideal edge detector. The core

concept of canny operator is to get the derivatives of

the first order of a gaussian function in any direction

to serve as a noise filter(Cai-Xia Deng, 2013). This

filter can be used to find the greatest value of the

INCOFT 2025 - International Conference on Futuristic Technology

local gradient. As a result, image edge can be

identified.

3.6 Proposed Algorithm

The accuracy of edge detection and efficient noise

reduction is the focus of this research. A gaussian

filter, which effectively removes noise has been

implemented for effective noise removal. The

benefits of each technique are combined in this

hybrid algorithm. The suggested approach as shown

in Figure 1 seeks to address the drawback of both

denoised gaussian image and the current edge

detection methods, Sobel and canny.

Step 1: Take any video as input, and if necessary,

add noise to it.

Step 2: To eliminate the noise from the video, apply

bilateral, gaussian and non-local filters.

Step 3: Use the PSNR and time as the two

evaluation measures to determine which method

performs the best.

Step 4: Apply the sober edge detector first to the

image and record the result as SE

Step 5: Next use the canny edge detector to examine

image and save the result as CE

Step 6: Create a hybrid of two outcomes as in

equation (10)

𝐻𝑦𝑏𝑟𝑖𝑑𝑖𝑚𝑎𝑔𝑒 = 𝑆𝐸+ 𝐶𝐸

(

)

4 RESULTS AND DISCUSSIONS

Three approaches have been utilized for denoising,

non local means denoising, gaussian filter and

bilateral filter. Videos of varying length have been

used as input for these algorithms. The original

image and outcome of the non-local means

denoising is compared in Figure 3(a), Figure 3(b)

contrasts the original image with gaussian filter,

Figure 3(c) compares the original image with

bilateral filter.

These algorithms have been evaluated using two

metrics: execution time and PSNR. We analyzed

videos of various lengths and execution time of all

three methods were recorded as shown in Table 1.

The bar graph in Fig 4 illustrates the same. Results

indicates that the Gaussian filter consistently

outperforms the other two algorithms across all

(a) Original Image Vs Non Local Means denoising

(b) Original Image Vs Gaussian Filter

Figure 3: Output of three different algorithms

videos. Additionally, it was observed that for videos

longer than two hours, non-local means denoising

requires over 24hrs to complete

Sobel-Canny Fusion for Effective Edge Detection in Gaussian Denoised Digital Images

Table 1: Comparison of three methods using execution

time

Video

Duration(min)

Non Local

Means(sec)

Gaussian

Filter(sec)

Bilateral

Filter(sec)

0.2 268 4 24

5 13602 137 573

10 23532 306 924

25 54765 639 6374

60 83589 1649 13668

142 167896 2849 22471

Figure 4: Comparative analysis of noise reduction

methods using Execution Time

In a similar manner, we have assessed the three

approaches using PSNR metrics, which evaluate the

video quality. Videos of various lengths were

analyzed and PSNR values for each of the three

methods are shown in Table2. The bar graph in Fig.

5 reflects the same information. It was observed that

using second metric PSNR, Gaussian filter surpasses

the other two techniques.

Table 2: Comparison of three methods using PSNR

Video

Duration(min)

Non Local

Means(dB)

Gaussian

Filter(dB)

Bilateral

Filter(dB)

5sec 30.15 31.22 30.85

10sec 30.07 31.49 30.31

30sec 31.44 33.31 32.28

1min 29.04 31.45 30.29

5min 28.11 29.44 28.67

10min 31.02 32.33 31.13

30min 29.65 30.19 29.89

45min 32.25 33.27 32.81

1 hour 29.32 31.69 29.81

Figure 5: Comparative analysis of noise reduction

methods using PSNR

The original image is displayed in Fig. 6(a). This

image is smoothed and the noise id removed when

gaussian filter in applied. After filtering smoothed

image is displayed in Fig. 6 (b). Edge detection is

more successful after filtering. The image after

applying the Sobel operator is displayed in Fig. 6(c).

It is evident that certain elements have been masked

by the image’s edges. The outcome of applying the

INCOFT 2025 - International Conference on Futuristic Technology

canny operator is displayed in Fig. 6(d). As the

canny operator exhibits the lower sensitivity to noise

than Sobel operator, it is evident that the edges are

clearer. However, some edges are taken up by the

Sobel operator rather than canny operator. In this

situation, hybrid strategy aids in producing the

effective outcomes. The algorithm’s final output is

depicted in Fig. 6(e). It incorporates the benefits of

the three elements.

Figure 6: Original Image and gaussian denoised image

Figure 7: Sobel edge detection and Canny edge detection

Figure 8: Proposed algorithms Hybrid image

5 CONCLUSION AND FUTURE

WORK

An essential stage in object extraction is edge

detection. An image uses less storage space when

edge detection is used. Edge detection is frequently

used for image segmentation since it can identify

objects and boundaries. Additionally edge detection

aids in identifying the patterns in an image by

extracting the key features. Hence it is essential to

obtain high quality outcomes from edge detection

methods. Individual operators cannot detect every

edge on their own, this paper proposed a hybrid

technique that leverages the benefits of both Sobel

and Canny Edge detection algorithms. Additionally,

the proposed algorithm incorporates a gaussian

filtering to eliminate any noise and to enhance edge

detection accuracy. The input image is smoothed

using this gaussian filtering. Additionally, this

gaussian filtering outperforms the other two

techniques. The suggested algorithm achieves a very

good edge detection effect and successfully

increases edge detection accuracy.

Various deep learning technique can be used to

perform this operation. The proposed algorithm is

suitable for short videos. However, as the video

duration increases, processing becomes significantly

time consuming. To address this, parallel processing

techniques can be utilized by distributing the

workload across multiple threads, thereby enhancing

the efficiency.

REFERENCES

Z. Liu, L. Yuan, X. Tang, M. Uyttendaele, and J. Sun

2014. Fast burst images denoising. ACM Transactions

on Graphics (TOG), 33:232.

M. Maggioni, G. Boracchi, A. Foi, and K. Egiazarian

2012. Video denoising, deblocking, and enhancement

through separable 4-d nonlocal spatiotemporal

transforms. TIP, 21:3952–3966

Mildenhall, J. T. Barron, J. Chen, D. Sharlet, R. Ng, and

R. Carroll 2018. Burst denoising with kernel

prediction networks. In CVPR

Antoni Buades, Bartomeu Coll, Jean-Michel Morel , Non-

Local Means Denoising, Image Processing On Line,

2011, pp. 208212.

https://doi.org/10.5201/ipol.2011.bcm nlm

Buades, B. Coll, J.M. Morel 2005, “A non local algorithm

for image denoising”, IEEE ComputerVision and

Pattern Recognition, Vol 2, pp: 60-65.

http://dx.doi.org/10.1109/CVPR.2005.38

Sobel-Canny Fusion for Effective Edge Detection in Gaussian Denoised Digital Images

Lingli Huang 2015, “Improved Non-Local Means

Algorithm for Image Denoising", Journal of Computer

and Communications, Vol.3 No.4

Sriani, Ilka Zufria, Mhd. Syahnan 2022, “Improved

Digital Image Quality Using the Gaussian Filter

Method”, International Journal of Information System

& Technology, Vol. 5, No. 5, pp. 556-563

Kumar and S. S. Sodhi 2020, "Comparative Analysis of

Gaussian Filter, Median Filter and Denoise

Autoenocoder," 27th International Conference on

Computing for Sustainable Global Development

(INDIACom), New Delhi, India, pp. 45-51, doi:

10.23919/INDIACom49435.2020.9083712

Priyanka, Rishabh Shukla, Laxmi Shrivastava 2020,”

Image Restoration of Image with Gaussian Filter”,

International Research Journal of Engineering and

Technology (IRJET) e-ISSN: 2395-0056, Volume: 07

Issue: 12

P. D. Patil and A. D. Kumbhar, "Bilateral Filter for Image

Denoising 2015. International Conference on Green

Computing and Internet of Things (ICGCIoT), Greater

Noida, India, pp. 299-302, doi:

10.1109/ICGCIoT.2015.7380477.

Bhonsle, Devanand & Chandra, Vivek & Sinha, Professor

G 2012. Medical Image Denoising Using Bilateral

Filter. International Journal of Image, Graphics and

Signal Processing. 4. 10.5815/ijigsp.2012.06.06.

Ghassan Mahmoud husien Amer and Dr. Ahmed

Mohamed Abushaala 2015, “Edge Detection

Methods”, IEEE.

A. Kalra and R. L. Chhokar 2016, "A Hybrid Approach

Using Sobel and Canny Operator for Digital Image

Edge Detection," International Conference on Micro-

Electronics and Telecommunication Engineering

(ICMETE), Ghaziabad, India, pp. 305-310, doi:

10.1109/ICMETE.2016.49.

Buades, B. Coll, J.M. Morel 2006, “A review of image

denoising methods, with a new one”, Mul-tiscale

Modeling and Simulation, Vol. 4 (2), pp: 490-530

http://dx.doi.org/10.1137/040616024

Buades, B. Coll, J.M. Morel 2005, “A non local algorithm

for image denoising”, IEEE ComputerVision and

Pattern Recognition, Vol 2, pp: 60-65,

http://dx.doi.org/10.1109/CVPR.2005.38

Gonzalez, R. C., and Richard, E. W. 2007, Image

processing. Digital image processing 2

Moslem Taghizadeh and Mahboobeh Hajipoor 2011, “A

Hybrid Algorithm for Segmentation of MRI Images

Based on Edge Detection”, International Conference

of Soft Computing and Pattern Recognition, pp. 107-

111

Cai-Xia Deng, Gui-Bin Wang and Xin-Rui Yang 2013,

“Image Edge Detection algorithm based on Improved

Canny Operator”, International Conference on

Wavelet Analysis and Pattern Recognition, pp. 168-

172.

M. G. Rao H. and G. S. Nagaraja 2019, "Noise Removal

Techniques and Quality analysis of X-ray Images," 1st

International Conference on Advanced Technologies

in Intelligent Control, Environment, Computing &

Communication Engineering (ICATIECE), Bangalore,

India, pp. 327-330, doi:

10.1109/ICATIECE45860.2019.9063843.

INCOFT 2025 - International Conference on Futuristic Technology