=Paper=
{{Paper
|id=Vol-1490/paper51
|storemode=property
|title=The Big Data methodology in computer vision systems
|pdfUrl=https://ceur-ws.org/Vol-1490/paper51.pdf
|volume=Vol-1490
}}
==The Big Data methodology in computer vision systems==
https://ceur-ws.org/Vol-1490/paper51.pdf
Data Mining and Big Data
The Big Data methodology in computer vision systems
Popov S.B.
Samara State Aerospace University,
Image Processing Systems Institute, Russian Academy of Sciences
Abstract. I consider the advantages of using the big data methodology in the
computer vision systems. It is noted that this solution provides transparent
increasing the functionality of CVS and improvement its quality, the formation
of new intellectual properties of the system based on the possibilities of a
posteriori multivariate iterative processing of stored video in the background.
The basic principles of such intelligent vision systems have been successfully
used to create a distributed vision system of railway tanks registration.
Keywords: big data, information technologies, computer vision system, image
processing, multimodal processing
Citation: Popov S.B. The Big Data methodology in computer vision systems.
Proceedings of Information Technology and Nanotechnology (ITNT-2015),
CEUR Workshop Proceedings, 2015; 1490: 420-425. DOI: 10.18287/1613-
0073-2015-1490-420-425
1. Introduction
The emergence and development of new approaches and technologies in
processing and analysis of big data has led to a shift in the methodology of formation
of new knowledge. Some researchers polemically declare the end of science, since the
scientific discoveries and insights can come from an exhaustive search of all possible
models of any scientific phenomenon with subsequent clustering. It certainly is
hyperbole, but there is a definite impact on modern information technologies.
Transfer of big data technology to another subject domain does not involve simple
copying a selected set of methods. It is important to understand how the basic
principles of a new methodology allow to move to a qualitatively new level of target
technology. Concrete solutions depend strongly on the specific application domain
and its level of hardware and algorithmic advances.
In this work we are primarily interested in the impact of data science approaches to
design solutions in the domain of computer vision.
2. The Big Data methodology
What is the main difference between the new approaches from the traditional data
processing technologies?
420
Information Technology and Nanotechnology (ITNT-2015)
Data Mining and Big Data Popov S.B. The Big Data Methodology in CVS...
New data, computational capabilities, and methods create opportunities and
challenges [1]:
─ Integrate statistics/machine learning to assess many models and calibrate them
against “all” relevant data.
─ Integrate data movement, management, workflow, and computation to accelerate
data-driven applications.
─ New computer facilities enable on-demand computing and high-speed analysis of
large quantities of data.
Many models. Traditional research approaches are based on the initial formation of
a certain model, then the accumulation and processing of data with subsequent
estimation of the parameters of such pre-formed model. In the process of discovery of
new knowledge big data technologies integrate simulation, experiment, and
informatics. This integrated research solves the problem of finding the model or a set
of models most appropriate to some experimental data. Moreover, this methodology is
aimed at finding the most interesting (i.e., unexpected and effective) and robust
models.
Data-driven applications. The relatively simple principles are at the core of big
data processing technology: “Divide and Conquer” – distributed architecture of data
storage system and total parallel processing at the lowest level; “Treat where Store” –
data are not moved during processing, the data processing tasks are delivered and run
on computing resources of distributed storage systems; “Data Are Forever” – data are
not deleted or modified during processing, the results simply are saved in situ.
New computer facilities. The widespread introduction of parallelism even in
commodity computers discovers new ability for developers of modern algorithms
allowing to implement a multi variance and competitive methods of processing. Easy
deployment of distributed systems allows transparently to increase intellectual
capabilities of target systems.
Seeing computer vision tasks, the most interesting is real-time big data processing.
Real-time big data analytics is an iterative process involving multiple tools and
systems. It’s helpful to divide the process into five phases [2]: data distillation, model
development, validation and deployment, real-time scoring, and model refresh. At
each phase, the terms “real time” and “big data” are context dependent.
1. Data distillation – This phase includes extracting data features, combining
different data sources, filtering for domains of interest, selecting relevant features and
outcomes for modeling, and exporting sets of distilled data to a local data store.
2. Model development – Processes in this phase include feature selection, sampling
and aggregation; variable transformation; model estimation; model refinement; and
model benchmarking. The goal at this phase is creating a predictive model that is
powerful, robust, comprehensible and implementable.
3. Validation and deployment – The goal at this phase is testing the model to make
sure that it works in the real world. The validation process involves re-extracting fresh
data, running it against the model, and comparing results with outcomes run on data
that’s been withheld as a validation set. If the model works, it can be deployed into a
production environment.
421
Information Technology and Nanotechnology (ITNT-2015)
Data Mining and Big Data Popov S.B. The Big Data Methodology in CVS...
4. Real-time scoring – At this phase the generated decisions are scoring by
consumers or by external control system.
5. Model refresh – Data is always changing, so there needs to be a way to refresh
the data and refresh the model built on the original data. The existing programs are
used to refresh the models in accordance with the newly accumulated data. It is also
recommended simple exploratory data analysis, along with periodic (weekly, daily, or
hourly) model refreshes.
Big Data methodology is the iterative process. Models should not be created once,
then deployed and left in place unchanged. Instead, through feedback, refinement and
redeployment, a model should continually adapt to conditions, allowing both the
model and the work behind it to provide value to the organization for as long as the
solution is needed [2].
3. Revised data processing technology in the computer vision
Turning to the problems of developing computer vision systems (CVS), it should
be noted that the traditional approach of using sequential processing steps performed
by the separate frames of video data largely exhausted the resources to increase
processing accuracy and the ability to adapt as conditions of surveillance change.
It is necessary to use some principles of big data technology. With regard to the
field of computer vision the most suitable solution are the following:
─ Total parallelization of the processing and distribution of data.
─ Multivariance and competitive data models and methods of their processing.
─ Continuous scoring of competitive solutions, both in the process of forming
solutions and to adapt to changing surveillance conditions.
The implementation of these solutions leads to the multimodal approach proposed
in this work. In this case, this term is broadly interpreted. First of all, the principle of
multimodal processing means the simultaneous use of a family of algorithms at the
critical stages, each of which is built on a fundamentally different approach. At the
same time, multimodality is a coordinated use of video data from different points of
view, with different spatial resolutions, a simultaneous use of multiple consecutive
frames and additional a priori information about objects of interest.
The results of such versatile processing are analyzed together. This achieves the
variability of approaches to solving complex problems, a significant increase in the
stability of technology in general with significant changes in observation conditions
and parameters of controlled or analyzed objects in video streams.
At a time when for a particular purpose computer vision requires a complex multi-
step process, to take a final decision on choosing the best one in a separate step is not
possible.
The effectiveness of a multimodal approach is best manifested in computer vision
applications in the presence of well-defined criteria for achieving the goal of
processing. Since in this case the final decision stage can be implemented as a
procedure of joint analysis of the whole set of variants generated by the multivariate
multistage process.
422
Information Technology and Nanotechnology (ITNT-2015)
Data Mining and Big Data Popov S.B. The Big Data Methodology in CVS...
A good example of the effective use of multimodal processing is the problem of
recognizing the numbers of cars trains. It is noted that similar systems for recognition
of train car number should be adapted for work under the following challenges:
presence of various digit outlines, variety of color combinations of digits and
background, distortions of numbers due to line of sight being non perpendicular to the
surface where the number is located (this is especially common for tank cars), various
kinds of contamination of the object’s surface, and the necessity to operate in both
artificial and natural lighting, with the latter accompanied by significant changes in
illumination conditions during the day.
The first design decision is the choice of distributed architecture for computer
vision system. Distributed modular software allows transparent mapping onto the
distributed computation system with a variable number of computers and makes it
possible to scale the system towards both a higher quantity of video streams being
processed and time reduce of video data processing.
Multimodal processing uses as the base the multistage recognition technology
developed for vision system of the railway train registration [3]. Revised technology
is as follows: at each stage of processing it is used the previously formed set of
algorithms that implement a completely different approaches; each algorithm from the
working set is applied to each frame of a video stream of the current wagon. More
than one result is obtained at all stages. The results lists are generated. All results are
ordered on the basis of relevant metrics. Thus a decision tree, corresponding to one
wagon video stream processing, is obtained. When forming the recognition result,
code protection of numbering system of train wagons and containers is used and it
takes into account the characteristics of validity of recognition formed for each digit.
These characteristics are accounted with weights that correspond to algorithms used
in previous steps of the process of obtaining a fragment of number. The final scoring
is provided by the operator who is responsible for confirmation or adjustment of the
final results within the system. It provides additional opportunities for improving the
operational quality of recognition technology. With correct result available, it makes
sense the posterior multivariate processing of archived video data in the background
iterative mode. In this way the quality of generated results may be estimated and
parameters of processing algorithms may be selected.
The using structured lighting [4] makes it possible to expand the spectrum of
algorithms of fast symbol localization by using the alternative method of delineation
of areas with no symbols, which uses the analysis of luminous line geometry of
structured lighting for recognition of structural elements of train wagons and
containers. Acquired data on the positions of the retaining bands on tank cars and
stiffening plates on train cars and containers may be used for the identification of the
type of the currently processed train car, tank, or container and therefore obtaining
additional information on typical car number locations for them. It is a clear
manifestation of the proposed multimodal processing principle.
Allocation of special computational capacity for data storage makes it possible to
extend the set of goals of posterior video data processing. This procedure makes it
possible to analyze and reveal additional information on the data located in the
system, retrieve specific frames of interest for further viewing, interpretation, etc., and
423
Information Technology and Nanotechnology (ITNT-2015)
Data Mining and Big Data Popov S.B. The Big Data Methodology in CVS...
generate an annotated list of detected abnormalities of the observation process [5],
ensuring fast access to them for the operator. The indicated opportunities make it
possible to use the system for security control. Further analysis of this information
allows CVS developers to improve processing and recognition algorithms.
The implementation of this approach involves the development of modifications
of the basic algorithms of fast localization of the alphanumeric information in
multiple video streams, enabling to form an ordered list of results; the formation of a
representative set of recognition algorithms; the implementation of tasks of analysis
of the obtained list and selection of the most reliable results.
The distributed architecture of the system ensures further increase in real time
video stream processing rate due to scalability of computations [6] within hardware–
software platforms of distributed multiprocessor computation systems, in particular
using CUDA technology within NVIDIA data processing acceleration hardware.
Distributed video data archive storage on several servers is based on the cascade
principle [7]: the working storage with which the video server interacts and the set of
archive storages, each one containing video data for a specific period of time. As the
working storage is filled, the data are transferred to a first level archive, where space
is cleared beforehand by transferring existing data to a second level archive. The same
procedure is performed throughout the whole cascade of archives. The most outdated
information is deleted at the bottom level.
The presence of the distributed archive storage needed for the implementation of
the first three phases of the process of real-time big data analytics and subsequent
model refresh.
5. Conclusion
The use of new technologies has a cumulative effect. Their implementation opens
new opportunities and challenges. The use of distributed software architecture in CVS
design considerably facilitates the transfer to advanced technologies based on cloud
computing. When complimented by an extended stock list of high resolution IP
cameras with possibility for wireless connection to the communication network, it
makes it possible to efficiently use three most promising technologies, which are now
showing an explosive growth of interest, namely computer vision, cloud computing,
and wireless data transfer. Integration of these three technologies significantly
expands the CVS application market due to a higher quality of services, a significant
reduction in time and complexity and, therefore, implementation costs, and
minimization of cost of ownership.
Acknowledgements
This work was partially supported by the Ministry of education and science of the
Russian Federation in the framework of the implementation of the Program of
increasing the competitiveness of SSAU among the world’s leading scientific and
educational centers for 2013-2020 years; by the Russian Foundation for Basic
Research grants (# 13-07-00997, # 13-07-12181, # 13-07-97002, # 15-29-07077); by
424
Information Technology and Nanotechnology (ITNT-2015)
Data Mining and Big Data Popov S.B. The Big Data Methodology in CVS...
the Presidium of the RAS program # 6 “Problems of creation of high-performance,
distributed and cloud systems and technologies” 2015.
References
1. Foster I. Taming Big Data: Accelerating Discovery via Outsourcing and Automation. Keynote
Lecture, the International Winter School on Big Data Tarragona. Spain, January 26-30, 2015.
2. Barlow M. Real-Time Big Data Analytics: Emerging Architecture. O’Reilly Media, 2013; 30
p.
3. Kazanskiy NL, Popov SB. The distributed vision system of the registration of the railway
train. Computer Optics, 2012; 36(3): 419-428. [in Russian]
4. Popov SB. The intellectual lighting for optical information-measuring systems. Proc. SPIE
9533, Optical Technologies for Telecommunications 2014, 2015; 95330 p.
5. Kazanskiy NL, Popov SB. Machine Vision System for Singularity Detection in
Monitoring the Long Process. Optical Memory and Neural Networks (Information Optics),
2010; 19(1): 23-30.
6. Kazanskiy NL, Popov SB. Distributed storage and parallel processing for large-size
optical images. Proc. SPIE 8410, Optical Technologies for Telecommunications 2011,
2012; 84100I.
7. Kazanskiy NL, Popov SB. Integrated Design Technology for Computer Vision Systems
in Railway Transportation. Pattern Recognition and Image Analysis, 2015; 25(2): 215-219.
425
Information Technology and Nanotechnology (ITNT-2015)