Decoding the dichotomy:

Traditional Image Processing

vs.
Deep Learning

Whitepaper

Highlights

 Comparison of Deep Learning and Traditional Image Processing

 Hybrid Approaches
 Guidelines for making a suitable choice
Photo by Pablo García Saldaña
on Unsplash

Note
For the context of this paper, the word “traditional image processing” shall be used to refer to a broader area of image processing which encompasses
domains of image processing, computer vision, and classical machine learning.
Whitepaper: Traditional Image Processing vs. Deep Learning

1. Introduction
Deep learning has certainly revolutionized traditional image In contrast to traditional image processing techniques, DL
processing. It has pushed the boundaries of Artificial helps achieve greater accuracy in tasks such as object
Intelligence to unlock potential opportunities across industry detection, image classification, Simultaneous Localization
verticals. and Mapping (SLAM), and semantic segmentation.

Figure 1: (a) Traditional Image Processing workflow vs. (b) Deep Learning workflow.

Several challenges that once seemed impossible to solve, are Since neural networks used in DL are trained rather than
now solved to a point where machines are performing better programmed, applications following this approach often
than humans. However, that does not mean that the require less fine-tuning and expert analysis. The availability of
traditional image processing techniques that have advanced a humongous amount of video data in today’s system
in the years before the rise of DL have been made obsolete. supports this cause. While CV algorithms tend to be more
domain-specific, DL, on the other hand, provides superior
This paper will analyze the benefits and drawbacks of each flexibility because CNN models and frameworks can be re-
approach. This paper aims to provide better clarity on the trained using a custom dataset for any use case.
subject which can help data scientists/ industries choose the
most suitable method depending on the task at hand.

2. Comparison of Deep Learning and

Traditional Image processing
2.1. Advantages of DL
Rapid advancement in DL and enhancements in device
capabilities including memory capacity, computing power,
power consumption, optics, and image sensor resolution have
accelerated the spread of vision-based applications along
Figure 2: Data vs Performance Comparison
with improved performance and cost-effectiveness.

Technology for the next decade, today.

www.hcltech.com Page 2 of 6
Whitepaper: Traditional Image Processing vs. Deep Learning

2.2. Advantages of traditional Image processing will behave outside the training environment. It also offers
flexibility to CV engineers to tweak their parameters to either
At times, deep learning is overkill as traditional image improve their algorithm to achieve better accuracy and
processing can often solve a given problem with greater performance or investigate their mistakes when the algorithm
accuracy and in fewer lines of code than DL. The features fails. Traditional image processing is preferred for edge
learned from a deep neural net are specific to the training computing too, owing to its delivery of high performance with
dataset which if not well constructed, probably won’t perform lower resource usage. This also makes traditional image
well for images different from the training set. In contrast, processing more popular for cloud-based applications where
algorithms like SIFT and even simple color thresholding and high-powered resources that are required for deep learning
pixel counting algorithms are not class-specific, that is, they applications are expensive.
are very general and perform the same for any image.

2.3. Hybrid Approaches

Hybrid approaches are an amalgamation of traditional image
processing and deep learning that present the best of both.
They are gaining importance owing to their ability to maintain
the right balance between mature and proven traditional
image processing algorithms and versatile and accurate deep
learning techniques.

Hybrid approaches have witnessed resounding success in

medical image processing. Doctors can generally diagnose if
a tumor is benign or malignant through mammal review, but
hybridizing DL and CV capabilities allows us to automate this
process and reduce the possibility of human error.

Figure 3: Example of image stitching They are notably efficient in high-performance systems that
require quick development. For example, an image
processing algorithm can competently perform face detection
Therefore, SIFT and other algorithms are often preferred for over the live feed from a security camera. These detections
applications such as 3D mesh reconstruction/ image-stitching can then be relayed to a DNN as the next stage for face
which do not need specific class knowledge. While the recognition.
solutions to these tasks can be attained by training huge
datasets, the vast research effort required for it is not feasible
for a closed application. Summarily, deciding on the most
suitable approach for a given computer vision problem, one
should consider practical feasibility.

A product classification problem can be considered as an

example. Supposing the problem aims to classify cans of food
on a conveyor belt into either vegetarian or non-vegetarian
distinguished by can color – green for veg. and red for non-
veg. While the problem can be solved using accurate DL
models generated by collecting sufficient training data, the
traditional image processing is a much-preferred alternative in
this scenario with its simple color thresholding technique. This Figure 4: Example of hybrid approach – face recognition
example also highlights the fact that DL often fails to
generalize the task at hand in the event of limited training This helps the DNN to work only on a small patch of the image
dataset leading to over-fitting. thereby, reducing the considerable amount of computing
resources and training effort that would otherwise have been
Manual tweaking of parameters of a model is a daunting task
required to process the entire frame.
since a DNN consists of parameters in the order of millions
inside it, each with complex interrelationships. As a result, DL Fusion can also help achieve better accuracy. One such
models are censured to be a black-box. On the contrary, classic example is document processing where traditional
traditional image processing offers complete transparency image processing techniques are used for pre-processing
and allows one a good estimate of how his/ her techniques tasks like noise reduction, skew detection/ correction, and

Technology for the next decade, today.

www.hcltech.com Page 3 of 6
Whitepaper: Traditional Image Processing vs. Deep Learning

localization of lines and words. This, when followed OCR

using deep techniques, yield better accuracy. ImageNet – 15 million images with 1000 object
categories/ classes
The blend of machine learning metrics and deep networks has
gained significance over the years, owing to the evidence that Open Images v6 – 9 million images, with 600 object
it results in better models. Hybrid vision processing categories/ classes
implementations have proved a performance advantage while
Microsoft Common Objects in Context (COCO) – 2.5
providing a 130x-1,000x reduction in multiply-accumulate
million images, 91 object categories
operations and about 10x improvement in frame rates
compared to a pure DL solution. Furthermore, the hybrid PASCAL VOC dataset – 500K images, 20 object
implementation requires significantly lower CPU resources categories
and about half of the memory bandwidth.
Another point to be noted is the difficulty faced by ML models
in dealing with priors. It simply refers to the fact that not
3. Guidelines for making a suitable everything can be learned from data which requires more
choice priors to be injected into the models. Applications associated
with 3D CV such as image-based 3D modeling depict such a
Andrew Ng, one of the famous AI practitioners once quoted: scenario. For optimal performance, such problems require
“The analogy to deep learning is that the rocket engine is the strong priors such as smoothness, silhouette, and illumination
deep learning models and the fuel is the huge amounts of data information. CV approaches can either work independently or
we can feed to these algorithms.” It can also be extrapolated be complemented and work with DL for the following
to say that the quality of fuel (data) also plays a prominent role applications: morphing/ blending, sharpening, optical
in the performance of the rocket (model accuracy). corrections/ transformations, calculating geometries,
segmentation, de-blurring.
Deep learning is certainly impressive and exciting, but it is not
suitable for every situation. There are certain circumstances The below guidelines summarizes the common attributes of
where deep learning is probably not the best solution. Popular each technology from the preceding discussions. These
conventions regard deep learning unsuitable for below use- guidelines also act as a handy tool for not only data scientists
cases: but novice developers and business people who do not
 Augmented/ virtual reality necessarily have a thorough understanding of the subject to
make better decisions.
 3D modeling
 Video stabilization Prefer Deep Learning when:
 Motion capture/ calculation  Huge training data available for making accurate
 Noise reduction decisions.
 Possess high-computing power (i.e., CPU, GPU, TPU,
 Image registration
etc.) - to allow intensive model training and good
 Stereo processing application performance.
 Data compression and coding  Uncertain about the positive feature-engineering outcome
(i.e., selecting the most suitable feature(s) yielding the
desired outcome), especially in unstructured media
DL triumphs at solving closed-end classification problems, (audio, text, images).
which aims at mapping a range of potential signals to a finite  Deployment restricted to high-performance devices (i.e.,
number of categories, given that there is sufficient training data unsuitable for embedded, micro-controllers).
available and the test set does not deviate too much from the  Less/ no domain expertise is available
training set. However, an aberration from these assumptions
leads to critical problems that DL often fails to acknowledge. Stick to traditional image processing when:

Popular object detection datasets provide stats about the data  Limited (annotated/ labeled) data available.
size that is required by deep learning for achieving good  Lack of high storage and computing power.
accuracy.  Cheaper solution desired
 Desire flexible deployment over a range of hardware.
 Good domain expertise present.

Technology for the next decade, today.

www.hcltech.com Page 4 of 6
Whitepaper: Traditional Image Processing vs. Deep Learning

The below table summarizes the comparison between deep learning and traditional image processing:

Traditional Image
Selection criteria Deep Learning
Processing
Training dataset Small Large

Computing power Low High

Feature engineering Required Unnecessary

Training time Short Long

Annotation time Short Long

Algorithm Transparency High Low

Domain expertise High Low

Priors (Assumptions) Few Many

DLLs Model files, DLLs
Proprietary material - Risk of exposure
(Risk – Negligible) (Risk – High)
Deployment flexibility High Low

Expenditure (BOM) Low High

Some typical applications of DL and traditional image processing are given below:

Traditional Image Processing Deep Learning

Image transformation ( Lens distortion Image classification (OCR and Handwritten

correction, view changes) character recognition)

Image Signal Processing (ISP) Object detection/ identification

Camera calibration Semantic segmentation

Industrial inspection – Defect detection Instance segmentation

Stereo image processing Image synthesis

Automatic panorama stitching Image colorization

3D data processing Image Super-resolution

Calculating geometries Scene understanding

Technology for the next decade, today.

www.hcltech.com Page 5 of 6
Whitepaper: Traditional Image Processing vs. Deep Learning

4. Conclusion 5. References
This paper presented arguments for why traditional image [1] N. O’Mahony, S. Campbell, A. Carvalho, S.
processing techniques are still very much relevant in this Harapanahalli, G. V. Hernandez, L. Krpalkova, D.
age of Deep Learning. Riordan, and J. Walsh, “Deep learning vs. traditional
computer vision,” in Science and Information
Deep learning can be seen as one of the tools to solve Conference. Springer,2019, pp. 128–144
image/ video processing problems along with other [2] Zdziarski Z. (2018, January 26), The Reasons Behind
programming languages, image processing techniques, the Growth of Computer Vision. Zbigatron. Retrieved
classifiers, and machine learning. Some typical applications from https://zbigatron.com/the-reasons-behind-the-
were also compared from the perspectives of traditional recent-growth-of-computer-vision/
image processing and DL and discussed how the former is
often a simpler alternative in tasks where DL is an overkill. [3] Shaikh F. (April 8, 2017), Deep Learning vs. Machine
Learning – the essential differences you need to know.
The paper also highlighted some areas which hint that Analytics Vidhya. Retrieved from
hybrid approaches are the way forward with their lower https://www.analyticsvidhya.com/blog/2017/04/compari
memory bandwidth and computation requirements, and high son-between-deep-learning-machine-learning/
performance and accuracy. We also witnessed some [4] AI vs Machine Learning vs Deep Learning: What’s the
examples where traditional image processing techniques difference?. Guru99. Retrieved from
improve DL performance (i.e., reduced training time and https://www.guru99.com/machine-learning-vs-deep-
lower processing power). learning.html

Some general guidelines were also discussed in the paper [5] Brownlee J. (2019, March 13). 9 Applications of Deep
which shall prove to be useful for beginners as well as Learning for Computer Vision.
experts in the computer vision domain to make better Machinelearningmastery. Retrieved from
decisions while choosing between deep learning and https://machinelearningmastery.com/applications-of-
traditional image processing to solve their problem. deep-learning-for-computer-vision/
[6] Tulluri R. (2017, November 29). Conventional
computer vision coupled with deep learning makes AI
better. Network World. Retrieved from
https://www.networkworld.com/article/3239146/convent
ional-computer-vision-coupled-with-deep-learning-
makes-ai-better.html

Technology for the next decade, today.

www.hcltech.com Page 6 of 6