智能系统参考库：计算机视觉在控制系统中的应用-4

需积分: 10 18 浏览量更新于2024-07-19 收藏 11.08MB PDF 举报

"《计算机视觉在控制系统中的应用-4》是由Margarita N. Favorskaya和Lakhmi C. Jain共同编辑的一本书，属于智能系统参考图书馆系列的第136卷。这本书共有324页，采用英文版。书中主要探讨了计算机视觉在实际生活中的应用，特别是在控制系统的领域。此系列由Janusz Kacprzyk和Lakhmi C. Jain担任主编，旨在发布一个易于访问和结构化的智能系统参考库，涵盖了理论、应用和设计方法。" 《计算机视觉在控制系统中的应用-4》深入研究了计算机视觉这一前沿技术如何与控制系统相结合，以解决实际问题。计算机视觉是通过模拟人类视觉感知来理解和解释现实世界图像的学科，它涉及图像处理、模式识别、机器学习等多个领域。该书可能涵盖了以下几个关键知识点： 1. 图像获取与处理：介绍了如何使用摄像头和其他传感器获取图像，并对图像进行预处理，包括去噪、增强、校正等步骤，以提高后续分析的准确性。 2. 特征提取：讲解了如何从图像中提取有意义的特征，如边缘、形状、纹理和颜色，这些特征对于识别物体、跟踪运动和理解场景至关重要。 3. 目标检测与识别：阐述了如何利用计算机视觉算法来检测和识别图像中的特定对象，包括基于深度学习的方法，如卷积神经网络（CNNs）。 4. 运动分析与跟踪：讨论了如何分析图像序列中的物体运动，以及如何追踪目标物体的位置和姿态，这对于视频监控、自动驾驶等领域非常关键。 5. 控制理论与应用：将计算机视觉与控制系统理论相结合，讲述了如何利用视觉信息来实时调整和优化控制策略，例如在机器人导航或自动化生产线上。 6. 实时性能与优化：探讨了在实际应用中如何实现高效、实时的计算机视觉算法，以满足控制系统的速度和精度要求。 7. 深度学习与大数据：介绍了深度学习在计算机视觉中的作用，特别是在处理大规模数据集时如何训练模型以提升识别性能。 8. 安全与隐私：讨论了在使用计算机视觉技术时如何保护个人隐私和系统安全，这是现代技术发展中的一个重要议题。 9. 应用案例：书中很可能包含了多个实际应用案例，展示了计算机视觉在自动化、制造业、交通、医疗等领域的具体应用。《计算机视觉在控制系统中的应用-4》是一本全面且深入探讨计算机视觉在控制领域应用的专业著作，对于从事相关领域的研究人员、工程师和学生来说，具有很高的参考价值。通过阅读本书，读者可以了解当前计算机视觉技术的最新进展，以及如何将这些技术有效地应用于控制系统的设计和实施中。

1.1 Introduction

The core of any control vision system is the algorithms for raw data processing in

such manner that it will be possible to obtain the non-evidence dependences in big

data volumes, essential improvement of visual data, or reliable recognition of the

objects of interest. Each task requires the special approaches in creation of inno-

vative algorithms, as well as the great experimental work before its implementation

in real devices. Sometimes, visual data are not enough for decision making. In these

cases, audio data can be attracted successfully. The spectrum of real life applica-

tions is very wide. Therefore, this book is an attempt to contribute in some spheres

of human activity including culture, robotics, human interactions, and medicine.

1.2 Chapters Included in the Book

Chapter 2 includes the issues of graphological analysis and identiﬁcation of

handwritten text s on the examples of the author’s calligraphy courtesy provided by

the Manuscript Department of the Institute of Russia n Literature (Pushkin’s House)

of the Russian Academy of Sciences [1]. The authors discuss the main challenges of

this problem due to great variety of historical texts, when even the samples of

writing of some letters can greatly vary depending on the speed of writing and the

writing tool used (goose quill, pen point, or pencil). Since a textual analyst follows

the own manual algorithm for handwritten text identiﬁcation, it is possible to design

the automated information storage and retrieval system as a useful software tool in

order to reduce such long process. Additionally to the existing methods and

algorithms of bitmap image transformation, two methods were developed.

Vectorization of bitmap images of handwritten texts builds the orientation ﬁeld of

image contours using the following steps: the pre-ﬁltering of the bitmap image,

construction of the orientation ﬁeld for this bitmap image, ﬁltering and extrapola-

tion of the direction ﬁeld, and searching for and tracking of image contours using

the created orientation ﬁeld. The core idea of this algorithm is to trace the contours

in the original image that provides information about the direction of a line or a

stroke near a given point. Method of vector dynamic parameterization permits to

obtain more information about letter images. This method try to restore the dynamic

information about the movement of a pen, when writin g a letter, including the

indication of the start and end points, movement direction , and the number of cyclic

letter outlines . Thus, the vector dynamic representation provides the ability to

visualize each letter in the form of a 3D image with x, y and t coordinates, where t is

a time. Also the reader can ﬁnd the detailed description of architecture of the

designed software tool based on the client-server technique [2]. The system inter-

acts with the database (aroun d 400 Mb memory), which is pre-ﬁlled with a col-

lection of different versions of the author’s calligraphy of the classics of Russian

2 L.C. Jain and M.N. Favorskaya

literature A.S. Pushkin and A.S. Griboyedov (images of individual letters and their

ligaments/ligatures).

Chapter 3 investigates the issues of perceptually tuned watermarking using

non-subsampled shearlet transform. The multiple classiﬁcation and criteria of the

watermarking techniques permitted to chose very promising transformation called

as digital shearlet transform and its modiﬁcations for embedding/extraction of

gray-scale or color watermarks into the host images [3]. The tampering attacks

categorized as the global and local ones distort a watermark; more several types of

attacks can be applied for the same watermark. This causes a necessity to develop

the robust watermarking algorithms, also considering the contents of the host image

and watermark [4]. Such algorithms ought to ﬁnd a decision in a contradictory

desire to embed the maximum data volume (an image payload) and a visibility. The

multiplicative watermarking techniques support a balance of the signal magnitude

(ﬁdelity) and the quality of the watermarked image (robustness) at an acceptable

level. The multi-scale decomposition analysis is provided by many wavelet-like

methods, including ridgelet, curvelet, brushlet, wedgelet, beamlet, contourlet,

bandelet, directionlet, and shearlet. The use of the digital shearlet transform pro-

duces a variety of sub-bands for inserting the secret data due to the shearlets’

multi-resolution proper ty. These sub-bands are correlated in the digital wavelet

transform but not correlated in the digital shearlet transform that makes a water-

marking more secure procedure. At the same time, the non-subsampled shearlet

transform is a fully shift-invarian t, multi-scale, and multi-directional expansion of

the shearlet transform . The proposed algorithm of a watermark embedding and

extraction using a watermarking scrambling via Arnold’s transform and paradigms

of a human visual system demonstrates good experimental results on 44 color and

monochrome images from the database Miscellaneous of University of Southern

California. The scaling, disproportionate scaling, rotation, translation, and cropping

attacks were simulated, and the impact of the attacks on a watermark was estimated

by bit error rate and the peak signal to noise ration metrics.

Chapter 4 conducts the issues of the Simultaneous Localization And Mapping

(SLAM) algorithm improvement for the task of the mobile robot simultaneous

localization and mapping, which consists of two parts: the evaluation of the tra-

jectory of the robot movement and evaluation the locations of landmarks, which

depend on the coordinates of the robot at the time of each measurement. [5]. The

authors consider the FastSLAM modiﬁcation using the nonlinear models throu gh

the ﬁrst order Taylor series expansion at the mean of the landmarks’ observations.

The probabilistic properties of the SLAM algorithm were considered. The proposed

algorithm computes a more accurate mean and uncertainty of the landmarks moving

nonlinearly. The method of calibration of Kinect-like cameras, depth map

restoration using a modiﬁed interpola tion technique, and ﬁltering the noise in the

RGB images are the main preprocessing methods. Additionally, the improved

resampling algorithm for the particle ﬁltering throu gh the adaptive thresholding

based on the data of the effective particle number evolution was develo ped. The

proposed algorithm runs in real time and shows good accuracy and robustness in

comparison with other modern SLAM systems. The average errors in calculating

1 Innovative Algorithms in Computer Vision 3

the displacement of the camera between successive frames using the adaptive

threshold values is about 24% less than in the case of the strict thresholding.

Chapter 5 implies the development of fast parallel algorithms for joined pro-

cessing of visual and audio information in the motion control systems of the mobile

robots [6]. The main goal is the path planning of a mobile robot in real time

implementation using three types of information obtained from the video camera,

microphones, and laser range ﬁnders [7]. Such algorithms ought to have the

appropriate precision and effectiveness in scenes with the moving objects, people,

and obstacles. Fast parallel algorithms are implemented using the multiprocessor

hardware structure (the well known NVIDIA Graphics Processing Unit

(GPU) processor) and software platform (Compute Uniﬁed Device Arc hitecture

(CUDA) platform). Many mathematical operations widely used in the images,

audio, and other signal processing, such as the matrix multiplication, Fast Fourier

transform, correlation and convolution, can be executed faster in a comparison to

the traditional Cen tral Processing Unit (CPU). During this research, three algo-

rithms were developed. The Algorithm 1 is based on the SLAM method and the

mobile robot audio visual perception and attention [8]. The SLAM method collects

and keeps the information for all previous executive locations for a decision making

about the following locations of a mobile robot. The Extended Kalman

Filter SLAM (EKF SLAM) was modiﬁed for parallel execution. Algorithm 2 and

Algorithm 3 are the developed fast parallel algorithm based on the mobile robot

audio perception and attention for a decision making in motion control system

managed by a speaking person [9]. They consider the current audio features

deﬁning an angle of sound source arrival or location of sound source. The Steered

Response Power (SRP) with PHAse Transform (PHAT) called as SRP-PHAT was

applied for this purpose. The Algorithm 1, Algorithm 2, and Algorithm 3

demonstrate clearly the advantages and real time execution of using NIVIDA GPU

and CUDA platform instead of CPU in the developed fast algorithm for a decision

making in motion control system of a mobile robot based on audio visual

information.

Chapter 6 promotes the indoor ambient audiovisual monitoring in the intelligent

meeting room with a goal to determine the time events, when the states of the

participants’ activities are changi ng [10]. The chapte r contains a description of

some intelligent meeting rooms equipped with audio and video recording equip-

ment [11]. However, the main discussion deals with the methods and algorithms of

image and speech analysis, such as the recognition and identiﬁcation of faces,

detection and tracking of participants, identiﬁcation of participants’ positions, voice

activity detection, localization and separation of sound sources, identiﬁcation of

speakers, and recognition of speech and non-speech acoustic events [12]. Also the

combined methods are possible, for example, the estimation of the position and

orientation of the speaker’s head and multimodal recogniti on of emotions. Such

audiovisual monitoring permits to control the events (a new participant enters the

meeting room, the presentation begins, and the audience is given the ﬂoor) taking

place in the meeting room. The proposed image processing includes the removal of

low-quality frames, illumination normalization, elimin ation of image blur, cleaning

4 L.C. Jain and M.N. Favorskaya

of digital noise, and automatic segmentation and recognition of the participants’

faces. In experimental section, one can ﬁnd a digital comparison of three basic

techniques for face recognition, such as the principal components analysis, linear

discriminant analysis, and local binary pattern analysis. The authors developed a

method of audiovisual recording of participant activity based on the multifunction

system of video monitoring and multichannel system of audio localization that

detects the position of the sound source evaluating the phase difference of signals

recorded by pairs of microphones from four arrays [13].

Chapter 7 contributes in the audience analys is systems [14] and gender recog-

nition systems [15] using the assessment of face image quality. This is a continu-

ation of investigations of these authors during last two decades. Several standards,

such as ISO/IEC 19794-5 and ICAO 9303, contain a description of parameters that

provide a decision of the image suitability in the automatic face recognition sys-

tems. All parameters are grouped into two classes the textural features (sharpness,

contrast and light intensity, compression ratio, and other distortions) and face

features (symmetry, position, rotation, eyes visibility, and the presence of glare or

shadows on the face). For such algorithms, an assessment of face image quality is a

cornerstone because a success of face veriﬁcation/identiﬁcation depends strongly

from the selected lucky frames. Three types of the quality metrics were considered,

such as the texture-based, symmetry-based, and universal metrics. The authors

developed the no-reference image quality assessment algorithms based on the

analysis of texture and the landmark points ’ symmetry in the facial images [16].

Note the original methodology of the experiments’ statement. Two sets were

formed called as Top1 with the single best image of a person and Top3 considering

three high quality images chosen by the experts and objective metrics. The pro-

posed algorithms were tested using standard LIVE and TID2013 image database.

The performance results show that the proposed algorithms are highly competitive

for audience analysis systems and, moreover, have very low computational com-

plexity making it well suited for real-time applications. Additionally, novel gender

classiﬁcation algorithm was suggested based on non-linear SVM classiﬁer. It

includes several steps, among which are the scale-invariant feature transform,

histogram of oriented gradients, Gabor ﬁlters, and pre-selection of blocks with the

corresponding positive experimental results.

Chapter 8 contains a description of real time eye blink detection method for

Android device controlling and its implementation using standard libraries like

OpenCV conversion procedures. The Human Computer Interaction (HCI) or

Mobile Computer Interaction (MCI) may be realized in different ways. The authors

chose the issue of the eye blinking detection that has a signiﬁcant role in the human

mobile interaction, computer interaction, driving safety, and health care. The

heuristic propositions concluded from the theory and practice of the HCI were used

during design. The eye tracking is the fastest non-invasive method of measuring

user attention [17]. The best deﬁnition of eye tracking can be the estimation of user

gaze’s direction, which is very difﬁcult problem especially in the real world. The

proposed method consists of ﬁve main parts including the mobile camera pro-

cessing, face detection, eye detection, eye tracking, and blink detection. At the

1 Innovative Algorithms in Computer Vision 5

initial step, a video captured from the front camera of any Android device is

converted to grayscale frame using OpenCV, and then stored in special mobile

folder to be used later for face detection step. For face detection, Haar classiﬁers are

used, while the eyes detection is based on the trained AdaBoost and Haar feature

algorithms. The corneal reﬂection and pupil center of the eyes have been used as the

most important eye features in the proposed method in order to track the movement

of the eyes. During eye blink detection step, each black color pixel in the grayscale

frame represented by 1, while each gray or white color pixel will be represented by

0. The median blur ﬁltering is applied in this step. The mobile activity is controlled

in case the eye blinks by several ways, such as the sending a text mess age, turning

on the alarm system, opening a web browser or making a phone call. The exper-

iments with different distances between a human face and phone and vario us

lighting conditions (indoor/outdoor) were conducted. The overall and detection

accuracy had reached 98%. For each frame, the average execution time was

12.30 ms that provides a real time execution.

Chapter 9 contai ns the rich experimental study of the medical image processing

and morphological analysis in urology and plastic surgery (hernioplasty). At pre-

sent, it is necessary not only to identify the localization of the calculus but also to

determine the density and conﬁguration of the calculus and evaluate the functional

state of the urinary tract above and below the obstruction. For this purpose, the

methodology based on a novel method for color coding of contour representation

obtained by the digital shearlet transform was developed. Medical images contain

the noise of different natur e. In order to improve medical images that can have a

high resolution the optimized in implementation algorithms of the most frequently

used ﬁlters, such as the mean ﬁlter, Gaussian ﬁlter, median ﬁlter, and 2D Cleaner

ﬁlter, were developed. The study of properties of ﬁlters ’ functioning permitted to

use all three type of parallelism, viz. the data, algorithmic, and functional paral-

lelism. A comparison of the optimized and ordinary implementations of noise

reduction ﬁlters shows great speed improvement of the optimized implementations

(around 3–20 times). The highest increase in the processing speed was achieved for

the median ﬁlter. Thus, for the small kernel 3  3 the acceleration was about 8

times and for the large kernel 11  11 is around 70 times. For the parallel

implementation, the OpenMP standard was used. The chapter includes the

pseudo-code of some program procedures. For contour representation, the simple

conventional methods, such as Roberts, Prewitt, Sobel, and more complex ones

(LoG and Canny methods) were tested. However, the main attention was paid to

use digital shearlet transform in order to obtain the best results in contour extraction

[18]. This approach was enforced by the novel color coding algorithm. It is based

on the color selection and density distribution of the isolines in an image corre-

sponds to the known technique of building elastic maps using the spatial data. As a

result, the image accuracy of estimates in urology and plastic surger y (hernioplasty)

was increased up 10–25% in averaged. In urology, the proposed color coding

method increased accuracy, especially in complex cases of multiple stones in the

kidney. In hernioplasty, the color coding allows to conduct more efﬁcient analysis

6 L.C. Jain and M.N. Favorskaya

剩余323页未读，继续阅读

huachang

粉丝: 2
资源: 10

智能系统参考库：计算机视觉在控制系统中的应用-4

化工仪表及自动化控制系统PPT学习教案：现场总线与计算机控制系统特点、组成与优势

Intel® Open Source Computer Vision Library - OpenCV 中级教程

"二阶多智能体系统家居设计中的追踪控制研究

Computer Vision in Control Systems-4: Real Life Applications

Computer Vision in Control Systems-3: Aerial and Satellite Image Processing

Advanced Topics on Computer Vision, Control and Robotics in Mechatronics

人在回路机器学习Human-in-the-Loop_Machine_Learning.pdf

Geometric Computing: for WT, Robot Vision, Learning, Control and Action

(The Springer International Series in Engineering and Computer S

Smart Transmission Grid Vision and Framework

最新资源