019 Course) NI Engineering (2 uter E Fourth Year of Comp' ( ete v) (410252 (®)) piscrete Mathematics (210241 1) Concepts. ital Image Processing shods for Image Enancement using sptta and Frequency Doman, Segmentation. | | To Understand Digi To Study Various Met To Learn Classification Techniques for image ject Recognition. To Understand Image Compression ‘and Obj | us Image Restoration Techniques «To Study Variot Image Processing Applications. + To Understand various Medical and Satellite be able to- red for Digital Image Processing can Method for Image Enhancement. ‘oncompletion ofthe course, student will con: Apply Relevant Mathematics Req coz: Apply Spectal and Frequency Dom age segmentation. C03: Apply algorithmic approaches for Im pression and Object Recognition. C04: Summarize the Concept of Image Com COS: Explore the Image Restoration Techniques. 06: Explore the Medical and Satelite Image Processing Applications Unit I: Introduction to Digital Image Processing 07 Hour, Introduction, Fundamental step In Digital Image Processing, Components, Elements of visual percepin Image Sensing and Acquisition, mage Sampling and Quantization, Relationships between pixels dferet Colour Models, image Types, Image File Formats, Component Labeling algorithm. Introduction to OpenCV too! to Open and Display Images using Python or Eclipse C/C++. (Refer Chapters 1, 2,3.and1!) Unit 0 1 Introducti Image Fi Correlatio Frequency Gaussian), Unit 1 1 Introducti Threshold Transform, Unit IV Ir Image Co Compressio quantizatid Object Rec Methods an Unit V_ In Introductio Blind-decon Unie Medical Im Analysis (In Satellite In Photograph Sensing, EaUnit Image Enhancement 08 Hours Introduction to Image Enhancement and its Importance, Types of Image Enhancement Spatial Domain Image Enhancement : Intensity Transformations, Contrast Stretching, Histogram Fqualization, Correlation and Convolution, Smoothing Filters, Sharpening Filters, Gradient and Laplacian Frequency Domain Image Enhancement : Low Pass filtering in Frequency Domain ((deal, Hutterworth, Gaussian), High Pass filer in Frequency Domain (Ideal, Butterworth, Gaussian) (Refer Chapters 4, 5 and 6) Unit I Image Segmentation and Analysis 08 Hours Introduction to Image Segmentation and its need. Classification of Image Segmentation Techniques Threshold Based Image Segmentation, Edge Based Segmentation, Edge Detection, Edge Linking, Hough Transform, Watershed Transform, Clustering Techniques, region approach, (Refer Chapter 7) Unit IV Image Compression and Object Recognition 06 Hours Image Compression: Introduction to Image Compression and its need, Classification of Image Compression Techniques- run-length coding, Shannon Fano coding, Huffman coding. Scalar and vector quantization, Compression Standards-)PEG/MPEG, Video compression Object Recognition : Introduction, Computer Vision, Tensor Methods in Computer Vision, Classifications ‘Methods and Algorithm, Object Detection and Tracking, Object Recognition. (Refer Chapters # and 9) Unit V_Image Restoration and Reconstruction 07 Hours Introduction, Model of Image degradation, Noise Models, Classification of image restoration techniques, Blind-deconvolution techniques, Lucy Richardson Filtering, Wiener Filtering. (Refer Chapter 10) Unit VI Medical and Satellite Image Processing 07 Hours Medical Image Processing : Introduction, Medical Image Enhancement, Segmentation, Medical Image ‘Analysis (Images of Brain MRI or Cardiac MRI or Breast Cancer) Satellite Image Processing : Concepts and Foundations of Remote Sensing, GPS, GIS, Elements of Photographic Systems, Basic Principles of Photogrammetry, Multispectral, Thermal, and Hyper spectral Sensing, Earth Resource Satellites Operating in the Optical Spectrum. (Refer Chapter 12) uuuets of inten —— see Human isa SE enti i EYE —— WH image 46 Nel 47 Hig 48 Hig Contrast Stretching = --——n—eem—— “Thresholding anv Grey Level Slicing (Intensity Slicing) omen mmmn—nn Bit Plane Sing. mma Applications of Bit Plane Slicing ewnonemmmmm— Dynamic Range Compression (Log Transformation). Power Law Transformation Spatial and Intensity Resolution... 631 63.1 63 634 6341 63.1 63.1 63.4 632W__ Image Processing a7 48 49 410 4a. Table of Contents — ey TT) Image Subtraction - en Neighbourhood Processing 461 Low Pass Pitering (Smoothin 462 Noise. 463 Low Pass Averaging Filtor (Smoothing) 464 Low Pass Median Filtering Highpass Ftering. 7 High-Boost Filtering. 48.1 Advantages of High Boot Fitering Zooming —— - 491° Replication ~ 492 Linear Interpolation Solved Examples on Neighbourhood Processing. Difference between Point Processing and Mask Processing. SS O00 eeu Chapter 5: Histogram Modelling 5-1 to 5-21 —_—_—_—_—_——— sa 52 53 54 55 Introduction st 5.1.1 Mean and Standard Deviation of Histogram. Linear Stretching Histogram Equalization ‘Additional Examples on Histogram Modelling. Difference between Histogram Equalization and Contrast Stretching —___. lo — Chapter 6: Image Enhancement in Frequency Domain Hiwsn 61 62 63 Introduction. ‘The Fourier Transform, Discrete Fourier Transform (DFT). 63.1 Properties of Discrete Fourier TransforM nee 63.1(A) The Separabiity Property. 63.1(8) Translation Property (Shifting. Property) 63.1(C) Periodicity and Conjugate Symmetry Property. 63.1(D) Rotation Property... tt 6.3.1(E) Distributive and Scaling Property enn 63.1(F) Average Value Property 63.1(G) Laplacian Property (Second Derivative) 63.1(H) Convolution Property... 63218 16 27 2 13 ime 792 193 794 710 Addit “ee Chapter 8: | —_— a3 Into a2 Redu 83 Error aad 832 4 Loss 841 842 843 a4 85 Shann 6 Lossy’ 861 742 462 pag Camps Op 463 asta) Applitions/advanages of COMPES OPES == — 87 (PEG2 sap sepmenin sing te Second Devan te PGT te comps ps1. alam of Gaussian OEP a9 Vector ge Liking ——— —_ B10 Data 783 ait Videos 762 Hough Transform.. 76.2(A) Applicatons/Advantages of Ho 781 782 783 784 Splitand Merge——- Image Segmentation based on Thresholding 794 Global Thresholding ou‘Table of Contents 78S, 737 Image Processing 4 7.9.2 Local (Adaptive) Thresholding —— —— 793° Optimum Thresholding...... 7.94 Watershed Algorithm.. 7.10 Additional Solved Examples. em Chapter 8: Image Transforms > 81 Introduction. a2 Redundant and Irrelevant Data. a3 ErrorCriteria 83.1 Objective Error Criteria, 832 Subjective Eror Criteria. 84 Lossless Compression Techniques 841 Dictionary Based Coding —. 842 Run Length Encoding (RLE) B43 Statistical Coding.. 844 Hulman Encoding. 85 Shannon-Fano Coding. 86 Lossy Compression... 861 Improved Grey Scale (105) Quantization 862 Transform Coding ()PEG Coding) 863 Joint Photographic Experts Group (JPEG). 87 JPEG 2000... 88 Comparison of Lossless and Lossy Compression, 89 Vector Quantization and Scalar Quantization 810 Data Redundancies 811 Video Compression Standard enn 812 Solved Examples. Chapter 9: Object Recognition 9109-7 91 Object RECOgMN ene 9:44 Pattern and Pattern Classes... 92 Classifiers 92.1 Minimum Distance Classifier. 922 Template Matching Classifier (Correlation based Classifier) 923 Classifier Performance... 924 — Bayes Classifier 93° Computer Vision nu10.1 Introduction — 10.2 Degradation Model . 10.3 Degradation Functions 103.1 Noise and Degra ise a \dation, vos latio Discrete Degradation Model 10S Inverse Filtering 105.1 Pseudo-Inverse Filtering. 106 — Wiener Filter 10.6.1 Drawback of Wiener Filters... 10.7 Power Spectrum Equalisation (PSE) 10.7.1 Blind Deconvolution... Noise Models RGB Colour Model. 1122 NTSC Colour Model... 1123 — YCbCr Colour Model... 11.2.4 CMY and CMYK Models... 11.2.5 HSI Colour Model meena 1Q Colour Model mene mm 11.2.6 Comparison of RGB and YI 113 Pseudo-Colouring— en "Application eeenemeneenennnnemm 12.2 Medical Image Processing amen 12.3 Satellite Image Processing aan 12.3.1 Remote Sensing Process..mm- 12.3.2 _ Passive and Active Sensing 12.3.3 Advantages of Remote Sensing am---—~ 12.3.4 Limitations of Remote Sensingewwm-n--vwrvnnsnnrn 124 Photogrammetric Imaging Devices mmm 42.5 Hyperspectral Sensing vwm-—Introduction to a Image Processing 1,1 Introduction Human beings are primarily visual creatures who depend on their eyes to gather information around them. Of the five senses that human beings have, sight is what we depend upon the most. Not many animals depend on their visual systems; the way human beings do. Bats use high frequency sound waves. They emit sound waves which reflect back when they encounter some obstruction, Cats have poor vision but an excellent sense of smell. Snakes locate prey by heat emission and fish have ‘organs that sense electrical fields. 1.2 What do we Mean by Image Processing ? What happens when we look at an object 7 — The eye records the scene and sends signals to the brain. These signals get processed in the brain and some meaningful information is obtained. Let us take a simple example ; when we see fire, we immediately identify it ‘as something hot. Two things have happened here. 1) The scene has been recorded by the eye. 2) The brain processed this scene and gave out a warning signal, This is image processing !!! — We start processing images from the day we are born. Hence image processing is an integral part of us and we continue to process images till the day we die. So even if this subject seems to be new, we have been subconsciously doing this, all these years. The human eye-brain mechanism represents the ultimate imaging system. Apart from our vision, we have another important trait that is common to all human beings. We like to store information, analyse it, discuss it with others and try to better it. This trait of ours is responsible for the rapid development of the human race, - Early human beings strove to record their world by carving crude diagrams on stone. All the drawings that we see in old caves is just that; storing images seen, trying to analyse them and discussing it with others in the tribe. Refer Fig. 1.2.1. — This art developed through the ages by way of materials and skill. By the mid-nineteenth century, photography was well established. Image processing that we study starts from this era. Fig. 1.2.1 — Though it was stated earlier that the human eye-brain. mechanism represents the ultimate imaging system, image processing as a subject involves processing images obtained by a camera. With the advent of computers, image processing as a subject grew rapidly. — Images from a camera are fed into a computer where algorithms are written to process these images. Here, ‘the camera replaces the human eye and the computer does the processing. — Hence image processing as an engineering subject is basically manipulation of images by a computer.the handle of the tea CHF All family pictures: phe are ‘2-dimensionsl. If this take a: simple example. er ea nec ane one OP in Fig. 1.3.2, We are all ind. Here the voltage IS e. This is a typical to locate a dot on the responding time. identity cards etC. on i let us ‘statement is not cleat = consider a voltage signal shown familiar with a signal of this It varying. with respect 0 tim 4-dimensional signal. If we want ave all we need to Know ists C0 1% ¥) Fig. 1.3.2 wehy images are 2-dimension Let us See image shown in Fig. 1.3.3, consider the Fig. 1.3.4 Fig. 13.3 to locate the dot shown, we need} rections (x and y) Fig. 134. je are 2-dimensional fin = inthis case position in two di images that we Se ‘A typical image Is TeP' tere Gry yadare the spatial coordinates andi Jove (tour in the case of colour image) 3 Hence grey level f varies with respect to che coordinates. resented as shown in Fi = Electromag the simult ‘These two together as without thi produces m electricity outwards fi meters per: = Although radiation is made sour ‘tungsten fil Lightis a ba the human | from 380 nm Most ofthe life are take range of the We must ni which are ca X-rays and ri The electror terms of was and the freqt Here cis the:\ W_imare Processing 1.4 _ The Electromagnetic Spectrum — The apparatus shown in Fig, 1.2.2 will work only if light is incident om th Introduction to image Processing object. What we call light is actually a very small section of the electromagnetic energy spectrum The entire spectrum Is shown in Fig. 1.4.1 The optical epectrum 200 900 990 455 492 tExvor Far Near|Vilet Blue Groen’ Utra-wolot Gamma rays Cosmie rays: xrays a“ 40°%10%10%10710°10"101010°10"" 1 19 10° 10" 10° 10" 10° 10” 10° 10° 10"°10""10" 877 Sor 622 ax ym =| 77 1600 6000 Rod| Noar Medium Far Infrarod Radio waves Audio frequencies 1 o0"10" inn) ‘1010 Fig. 14.1 = PBlectromagnetic energy, as the name suggests, exists in the simultaneous form of electricity and magnetism. ‘These two forms of the same energy are transmitted together as electromagnetic radiation. One cannot exist without the other. A flow of electric current always produces magnetism, and magnetism is used to produce electricity. Electromagnetic radiation is propagated outwards from its source at a velocity of 300,00,0000 ‘meters per second (3 x 108 m/sec). = Although our natural source of electromagnetic radiation is the sun, there are also a number of man- made sources which, among many others, include tungsten filament lamps, gas discharge lamps and lasers. Lightis a band of electromagnetic radiation mediated by the human eye and is limited to a spectrum extending from 380 nm to [Link]. = Most of the images that we encounter in our day to day life are taken from cameras which are sensitive to this range of the electromagnetic spectrum (380 - 760 nm), We must not forget though, that there are cameras which are capable of detecting infrared, ultraviolet light, rays and radio waves too. — The electromagnetic spectrum can be expressed in terms of wavelength and frequency. The wavelength (1) and the frequency (v) are related by the expression c hike (144) Here cis the speed of light =3 x 10® m/sec Ex. 1.4.1 : Calculate the frequency of oscilation of green light. ‘Soin. : It has been known that green light has a wavelength of approximately 500 nm (500 x 10-%m) Its frequency of oscillations can be calculated using Equation (1.4.1). Ae 3x10 m/sec 500x10-?m vy = 6x104Hz Le, the frequency of green light is 600,000,000,000,000 eydes/sec! Hence it is more convenient to discuss electromagnetic radiation in terms of wavelengths (nm) rather than frequencies (H2). 1.5 _ Units of Intensity dle Iaput escpon)- Output (nage) Snputervision > Input Image) - Output (Description) a ?Vr stig dict ferenc: raphics -visoe pmate Image Processit 19 Introduction to image “The material provided in this chapter is primarily basic information which would be required in subsequent discussions. ur study of the human visual system, though not exhaustive, provides a basic idea of the capabilities of the eye In perceiving pictorial information. In this chapter, preliminary concepts of digital image processing are presented. Difference between one-dimensional ‘and two-dimenstonal signals is explained, Topics such as electromagnetic spectrum and inverse square law are discussed ‘with examples, Elements ofthe human visual system are presented, Basic anatomy of the human eye is explained with afew Iustrations, Perceptual characteristics such as brightness adaptation and logarithmic response to incident intensity in form of Weber's ratio are also introduced. “The concepts explained here willbe found useful in understanding image processing algorithms in subsequent, “This chapter forms the fundamental base required to understand image processing21 Introduction _______—~ Basic elements of an image processing system = Digital image processing is basically modification of images on a computer. The basic components of an image processing sysiem are shown below. (1) Image Acquisition. (2) Image Storage. (3) image Processing. (A) Display. (5) Transmission (if require). We shall discuss each one in detail Fig. 20.1 1) Image Acquisition : — Image acquisition is the first step in any image Processing system. The general aim of image acquisition 4s to transform an optical image (real world data) into an array of numerical data which could be later ‘manipulated on a computer, ~ Image acquistion is achleved by suitable cameras, We = meras for diferent applications. If we need an Xray image, we use a camera (Sin) that ls sensitive ta Xrays. If we want an plasty ‘se cameras which are sensitive to nf Image : sensing and Acquisition For normal images (Farnlly pictures etc,) we use came, which are sensitive tothe visual spectrum, In this bo. ine shall diseuss cameras (Sensors) which are sensi, only tothe visual range. photovoltaic devices : Photovoltaic devices consis semiconductor junctions. They are solid state arra,, ‘composed of discrete silicon imaging elements known 3, Photoses. Photovoltaic devices give a voltage outpy signal thatis proportional to the intensity ofthe incicen, ight No external bias is required as was in the case « photoconductive devices. = The technology used in solid-state imaging sensors is based principally on charge-coupled devices, common'y mown as Charged Coupled Devices CCDs. Hence the Imaging sensors are called CCD sensors. = The solid state array (CCD) can be arranged in two different configurations. (ba) Linearray CCD. (b2) Area array CCD. (b4) Line Arrays: = The line array represents the simplest form of Cco imager and has been employed since the early 1970s Line arrays consist of a one-dimensional array 0 Fig. 2.1.2 ~ A single line of CCD pixels are clocked out into the Parallel output register as shown in Fig. 2.1.2. The amplifier outputs a voltage signal proportional to th¢ contents of the row of photosites, One thing to note | that line array CCD scans only one line (hence is on*” W imager = In order array CC moving t activity. u = Thistec that you cafe). A element (62) An The pro line, Te mechan = Area dimens to Inve bbe ded possi SS Sl| in two of cco y 1970s array of {into the 1.2. The jal to the to note b> ce Is one- Image Processing {In order to produce a two-dimensional image, the line array CCD imager has to be used as a scanning device by ‘moving this array over the object by some mechanical activity. CCD (AGB) tne array Fig. 2.13 ~ This technique is used in flat bed scanners (the scanners that you come across in your laboratory oF in a cyber “cafe), A line array CCD can have anything from a few elements to upto 6000 or more. (2) Area Arrays: ~The problem with line arrays is that it scans only one line. To get a two-dimensional image, we need to ‘mechanically move the array over the entire image. SIRE SSS SS SQ Fig. 21.4 ‘Area arrays or matrix arrays consist of a two- dimensional array of photosites. They make it possible to investigate static real world scenes without any mechanical scanning, Thus much more information can bbe deduced from a single realtime glance than would be possible with line arrays, Sensing and Acquisition ‘Area arrays can be seen i the digital cameras that ‘we Use for video imaging. The area arrays are more versatile than the line arrays, but there Is a price to be ald for this. Area arrays are higher on cost and complexity — Area sensors come in different ranges. Le. 256 x 286, 490 x 380, 640 x 480, 780 x 575, CCD arrays are typically packaged as TV cameras. A. significant advantage of solid state array sensors is that they can be shuttered at very high speeds (1/10,000 secs), This makes them ideal for applications in which freezing ‘motion is required. 2) Image Storage: ~ All video signals are essentially in analog form 2 electrical signals convey luminance and colour with continuously variable voltage. ‘The cameras are interfaced to a computer where the processing algorithms ae written. — This is done by a frame grabber card. Usually a frame grabber card is a printed circuit board (PCB) fitted to ‘the host computer with its analog entrance port matching the impedance of the incoming video signal. ‘The A/D converter translates the video signals into dligal values and a digital image is constructed, Frame grabber cards usually have a A/D card with resolution of 8 - 12-bits (256 to 4096 gray levels), Hence a frame srabber card isan interface between the camera and the computer. — Frame grabber card has a block of memory, separate from the computers own memory, large enough to hold any image. This is known a5 the frame bufer memory.‘represents black while 1 represents white. It isa black ‘and white image in the strictest sense. These images are ‘also called bit mapped images. In such images, we have ‘only black and white pixels and no other shades of grey. Refer Fig 223. nage Sensing and Acc ge meeen me bees tive Be hat 2 Wars an be generated by mixing the three prim, een and Blue, in proper propor frau sages rch pixel Ix composed of RGB va mia each of these colours require @-its (one by) fa representation. Hence each pixels represented debits R(B-Dits}, (E-D!S), BEES) ‘A 2h-it colour image supports 16, 777, 216 di ‘combination of colours. Colour images can be easily converted to grey sc ffere Images using the equation x = 020R+0596+0.11B 22 = Ameasier formula that could achieve similar results s ReG+B x5 (222 MATLAB code for converting a colour image to » 2D) Grey Scale image ere each pas is usually sored as a byte (8-its) ed pa cas gg 8) ine Cen wale eae 6 he same mae hare back white and various shades of grey present In image Refer Fig. 2.24. WH image gure(i) imshow(u figure(2) imshow(w figure(3) imshowru Matlab 4) Half Te = itis ot better grey les printer level di can onl ground bilevel = You ha (hopeh But if basicall — Even ¢ Cinelud a white of see iMusion called!Prima portio 3B valu byte) § altferes rey sca (221 sults'is (222) toa gr nage %% WH image Processing 26 Image Sensing and Acquisition figure(1) ~The human eye integrates the scene that it sees imshow(uint 8(im)) Consider a simple example. Consider two squares of say figure(2) 0.03 x 0.03 sqinch. One of these squares contains 3 Jot imshow(uint 8(new)) of black dots while the other square contains fewer figure(3) black dot fmishow(uintB(new!)) When we look at these squares from a distance, the two squares give us a perception of 2 different grey levels Matlab has an inbuilt command for conversion ngb2Zaray. ‘This integration property of the eye is the basis for half, 4) Half Toning toning. In this, we take a matrix of a fixed size and depending on the grey level required, we fill the matrix with black pixels = It is obvious that a grey scale Image definitely looks better than the monochrome image as it utilizes more grey levels. But there is a problem in hand. Most of the printers that we use (inkjet, lasers, dot matrix) are all bi level devices, Le, they have only a black cartridge and can only produce two levels (black on a white back- ground). In fact, most of the printing jobs are done using bilevel devices. Let us take an example. = Consider a 3x 3 matrix This matrix can generate an fllusion of 10 different grey levels when viewed from a distance. = You have all read newspapers at some point of time (hopefully). The images do look like grey level images. But if you look closely, all the images generated are basically using black colour. Refer Fig. 225. Fig. 225 — Even the images that you see in most of the books (including this one) are generated using black colour on ‘a white background, In spite of this we do get an illusion of seeing grey levels. The technique to achieve an iMlusion of grey levels from only black and white levels is called half-toning.Sensing and Aca,on Sampling 3a Introduction. = We know that an image is basically a 2-dimensional representation of the 3-dimensional world. We have also studied that images can be acquired using a Vidicon for a CCD camera or using scanners. The basic requirement for image processing is that {mages obtained be in the digital format. For example, one ‘cannot work with or process photographs on identity cards unless he/she scans itusing a scanner. ‘The scanner digitizes the photograph and stores tt on the hard disk of the computer, Once this is done, one can ‘use image processing techniques to modify the image as per requirement. Ina Vidicon too, the output which ts in ‘analog form needs to be digitized in order to work with the images. To cut a Jong story short, to perform image processing, we need to have the images on the ‘computer. This will only be possible when we digitize the analog pictures. Now that we have understood the importance of digitization, let us See what this term actually means. ‘The process of digitization involves two steps Fig. 3.1.1: Steps of process of digitization In other words, Digitization ‘Sampling + Quantization. We have had exposure to these terms in the lower semesters in subjects like Principles of Communication Engineering, Signals and Systems and Signal Processing which dealt with I-dimensional signals. Let us take a brief: look at these concepts and move ahead to the 2-dimensional domain. a2 and Quantization Sampling and Quantization ‘The Sampling process converts a continuous time domain signal into a discrete signal which [Link] at specific instances of time. Sampling depends on the ‘Sampling frequency of the Analog to digital converter ‘The values obtained by sampling a continuous function usually comprise of an infinite set of real numbers ranging from a minimum to a maximum depending upon the sensors calibration. These values must be represented by a finite number of bits usually used by a computer to store oF process any data, In practice, the sampled signal values are represented by a finite set of integer values. This is known as quantization. Rounding of a number {s a simple example of quantization. With these concepts of sampling and quantization, now need to understand what these terms mean when ‘we look at an image on the computer monitor. ‘Higher the spatial resolution ofthe image, greater is the sampling rate Le. lower is the image area Ax Ay represented by each sampled point. Similarly, higher the grey level resolution (tonal resolution) more are the ‘number of quantized levels. Hence spatial resolution gives us an indication of the sampling while grey level resolution (tonal resolution) gives us an indication of the quantization. i. Spatial resolution — —> Sampling * Greylevel resolution —+ Quantization ‘We have already stated that an image can be considered as a 2-D array. Image f(xy) is arranged in the form of NxMarray 10,0) 100, 1) neerrrseienerens QM .0) (1,1), na bie 7 tiey=| 20) 12,1) M=1) 191-10)... AMA — Hence every image that is seen on actually this matrix. Bach element a pixel, Never forget this.1H inage Processing ee an image © sty a mat sce Sa ples, rence more the poeels, mare the $I ‘ en era hence bau the spa resolution is known asthe grey level mber of bits, better isthe tonal iy of picture depends om tonal Whenever We computer it sa pixels and each pixel sampling ‘The value of each pixel similarly higher the nut quality, Hence the quali and spatial resolution. xis only ones and zeros. Hence The computer understa tee ted in terms of these grey levels need to be represen zeros and ones. If we have two bits to represent the grey Levels, only + diferent grey levels (2?) canbe identified viz. 00,01 10, 11, where 00 is black, 11 is white and the other two are diferent shades of grey. ‘simiary, fe have 8 bits to represent the grey Teves, ‘we wil have 256 grey level (24) Henee more the bits, more are the grey levels and better i the tonal arity (quantitation). The total size ofthe image is NxM x, where mis the number ofits used Consider an Image in Fig, 321(a} We plot the pixel values of only the first row of this image, Tis is shown inFig 32.100) ‘The x-axis isthe number of samples or pixels in the first row (sampling) while the y-axis isthe gey evel or value ‘ofeach sample (quantization). ‘Now, comes the obvious question. As we kaow, more the samples and the bits, better & the image. What then shouldbe the dea values of sampling and quantization, Sampling and q, ) Fig. 2.2.1: Concept of sampling and quantization = This answer will vary from image to image. Give: isa table of sampling and the quantization value sampling and the quantization increase, the num\x bits required to store the image increases tremendo.. = The clarity increases, but storage space req Increases too, We hence need to get a trade-of! bess the two, For simplicity, we consider a square mai size NN. ‘Table 3.2.1 : Number of storage bits for various values of N and m - att z , 1 2 3 Se eas ag 2 32 | 19% a zo | 3072 | 4098 520 644 7,168 eo | 406 | aise | 19, 28 | 16304 | 20480 | 24576 | 2672 2.768 18 | 16305 S276 | 49152 | 65536 | siz90 | 98304 | 114688 | 1310 286 | 65536 18 85 181072 | 196608 | 262144 | 327.680 | 393216 | 459, y752 | 524,28 512 | 262,148 | 524, ;. = 288 | 786432 | 1,048,576 | 1,310,720 1,572,864 | 1,835,008 | 2,097.1° + | 1048576 | 20 :, ——_ 2ov7ase | 3145728 | 4194304 | szazna0 | 6201456 7,340,032 | 8,388, ‘es of reducing quantaton levels and reducing spatial eslation| - resol Separately. we WD image Pi MATLAB code f 96% Effects ofr clearall de aimread(‘zebr a=double(a); a=a44; bemax{max(a)) Jinput “how m j-b/(2"i); 696 Fefloor(a/(+1) fgure(1) imshowfuint8( fgure(2) imshow(uint( (©) image Comparing | takes place as w MATLAB code | 96% Down sam %% To seethe samples %% clear all309,608 wane Fin ¢ Processing 43 Samp) MATLAB code for reducing quantization levels 96% Effects of reducing the quantization values %% clearall imread{‘zebratif) image Jinput (how many bits do you want 124 8°); j-b/(2*i}; %% since total number of bits is equal to 2°1 Fefloor(a/(+1)) Fis(F1255)/max(max(F)); figure(1) imshow(uint8(a)) figure(2) imshow(uint8(F1)) % normalizing % (©) Image using 2-bits Fig, 3.2.2 (4) Image using 1-bit Comparing the images, we see that “false contouring” takes place as we reduce the number of grey levels. MATLAB code for reducing spatial resolution %9% Down sampling ‘%% To see the effects of reducing the number of samples 6% lear all, de a= imread(‘deepatif) end = 1; 96% This needs to be done else the value of goes on increasing %% fete; end figure(1),imshow(a) figure(2)mshow(c) figure(3), imagese(a),colormap(gray) fAgure(4) imagese(¢),colormap(gray) (©) Down sampled image displayed after zooming to ‘match the size of the original image Fig, 3.2.3 ~ Its clear from the images that the resolution reduces as ‘umber of samples reduce. To compare and understand the actual effects, we plot them together. To make sure that they appear to be of the same size, we upsample the second image, Comparing Fig, 32:3(a) and Fig. 3.2.3(c) we see that the second image has a “Checker hoard’ pattern due to the reduction of samples weeines Processing Tsopreference Curves — an Nand m in Fr ave seen the eects of reducing Wand ~ we hae 20 ed nko et ‘ feomplingresoton) and ™ abe of N restos orimages _ in 1965 attempted = hn ery iy by TS. Huang it 1965 sem erientally the ef 8 sale ‘simultaneously. There, vec by varying W and 1 : 2 shown to a group of fiterene types of images were ai not have Ht people. The st image wasthat whi oft for example a wom face = The seca image was one which had intermediate mount af information for example 2 small OUP Sanding together and the hd mage wa 86 whlch had a lot of detals example an image af 2 crow these images, Nand m were varied. Observers wert then asked to rank them according to thei subjective quality, These results were summarised in the form of Isopreference curves inthe N-m plane ~ Points ying on an isopreference curve carespended fo image of ecual subjective quay From the isopeference curves Huang concided that images wit large amount of details require few grey levels Since the isopreteence curve of the crowed & near verti ft means that fora fixed vale of N, the image i nearly independent of @ o) Fog. 33. cont. Sampling an Group Fig. 33.1 Physical Resolution = Bynow we know thata digital image, or image fo is composed of diserete pixels. These pixels are a {in a row and column fashion to form a rect. picture area = Clearly, the total number of pixels in an imay: function of size of the image and the number 0 per unit area (example: inch) in the horizontal as the vertical direction, The number of pixels ps Tength is referred to as the resolution of the disp! device (most of the deskjet printers have 670 dj 670 dots per inch). ‘Thusa3%2 inch image at a resolution of 300 pics {Inch would have a total of 540,000 pixels!! ‘In most books as well as in this book, image size is! 4 the total number of pixels in the horizontal die: ‘mes the total number of pixels in the vertical direct (ecample : 128 x 128, 512 x 512, 640 x 480 Although this convention makes it relatively s'™ {eget ge the total numberof pies ina im not specify the physical size of the image ° tons defined inthe paragraph above = Image Proc A 640 x 48¢ S inches wher On the other 12 Inch whe Inch — Another term ratio, ~The ratio of t unit length 0 aspect ratio. image have th Aspect Ex. 34.4 : Comp when printed by a Soln. : Since we have the mages Ex 842: i we that is 600 pixel original image, w image? Sol We know Aspects For the original in Now for the ratio buta width. He Hence the resSW image Processing = A 640 x 480 image would measure 666 inches by ‘inches when displayed or printed at 96 pixels per inch On the other hand, it would measure only 1.6 inch by 12 Inch when displayed or printed at 400 pixels per inch. Sampling and Quantization Ex. 3.4.3 : How much storage capacity is required to store lan image with size of 1024 x 768 and 256 gray levels ? Soin. : Storage capacity required = AxBxC Here AXxB = Physical size of image = 1024768 = Number of bits required to get 256 gray levels, we must have 8 bits (2* = 256) Storage capacity required = 1024 x 768 x 8 = 6291456 bits 2 ' | : Ex. 3.4.4 : A common measure of transmission for digital data is the baud rate, defined as the number of bits transmitted per second. Transmission is accomplished in packets consisting of a start bit, a byle (8-bits) of information and a stop bit. (@) How many minutes would it take to transmit a 1024 x 1024 image with 256 grey levels if we use & '56 k baud modem ? (b) What would be the time required if we use a 750 k band transmission line ? x Fig. 342, — Another term that we need to understand isthe Aspect ratio, ~The ratio ofthe image's width to its height, measured in unit length or number of pixels Is referred to as its ‘aspect ratio. Both, a 3 x 3 inch image and a 128 x 128 | Soln.: image have the same aspect ratio of 1 (2)Since we have 256 grey levels, we need 8-bits for Aspectratio = representing each pixel ‘Along with these @-bits, we also have a start bit. Ex 9.4.1 : Compute the physical size of a 640 x 480 image | Stop bit — ‘whe printed by a printer at 240 pixole per inch, al Soln.: ‘Since we have 240 pixels per inch, the physical size of + weap ofp i Sl. Ex. 3.4.2: I we want fo resize @ 1024 x 768 image to one per w | that is 600 pile wide wih the same aspect ratio as the | tn splay ‘orginal image, what should be the height of the resized | (b)! api - ie ‘ati but a width of 600 pixels, 7 NOWMt = Fepectraa * 133 “454 ae") Hence the resized image will be 600x451. ninege Process wim So jred mean 6S as re ievtresrea mean onaparsse © WSC CsI Image sie © S12xS12K20 imagesize = 2621640 512x512%10 acon = Osean vetaken = 27305sec= 45min rhis caper desis with converting & seame O88 onunvous inte and space ino an image that ca BE miputes. The teehnique of converting & processed by 00 is called digitization ontinuous signal Intw 4 discrete one I a is explained here. Digitisation comprises of fo STEPS ‘2 sampling and quantization. We study the effects of reducing sampling and _quariation onthe image chapter. Vist importance based oF Sampling an ral resolution is also 4 image #5 processing a ing 00 8 matrix i ii the image pt Spatial resolution and red using MATLAB code yp this conce also explaing at a2 as as as rive equation f te convcuon tec vain sampling and quantization. the fects of reducing samph Sm plain quantization. Explain nopreterence curves Explain non-unitorm sampling. compute the physical size of a 480% 201 ‘won printed by a printer at 320 dP. 4.1 __ Introduce’ — Image enhancem processing. As th original image is more sultable tha fe, the Image is er = Image enhancem technique. By sub varies from pers technique used to ‘person, but the foranother, = It is also import enhancement is a any extra inform Improves the st ‘working with the — Image enhanceme 1) The spatial do 2) The frequency — Lets start with e domain, discuss t ‘enhancement an frequency domal Enhancement in FImage Enhancement in Spatial Domain 4.1 _ Introduction = Image enhancement Is one of the first steps in image processing. As the name suggests, in this technique, the ‘original image is pracessed so that the resultant image is ‘more suitable than the original for specific applications {the image is enhanced. = Image enhancement is a purely subjective processing technique. By subjective we mean that the desired result ‘varies from person to person, An image enhancement technique used to process images might be excellent for ‘a person, but the same result might not be good enough for another. = It fs also important to know at the outset that image — ‘The term spatial domain means working in the giver space n this ease, the image. It implies working with the pixel values or in other words, working directly with the raw data. Let f(x, y) be the original image where Fis the grey level value and (x, y) are the image coordinates. For a B-bit — image, Fcan take values from 0 - 255 where A black, 255 represeuts white and all the values represent shades of grey. In a image of size 256 x 256, x « oP ie tolemage Processing — tn point processing, we work with single pixels ie. Tis 1.x 1 operator. It means that the new value fx, 9) depends on the operator T and the present f(x, y). This statement will be clear as we start giving some examples = Some ofthe common examples of point processing are hancement in S) WH tmage Pro: MATLAB progr 90% MATLAB ¢ Modified dlearall grey level de * imread (‘sa 10 (125 figure(t) ‘Original grey lovelr colormap (gray Fig. 43-100) imagesc(a) figure(2) eolormap {gray imagesc(b) 43.1 Digital Negative Digital negatives are useful In a lot of applicai, ‘common example of digital negative is the display, fan K-ray image. As the name suggests, nega ‘means inverting the grey levels 4e. Black in the ong {mage will now look white and vice versa. The Fix ¢ {the digital negative transformation for a 8-bit imy. ‘Onginal grey levelr 255 Fig. 4.3.2 ~ Before we proceed to explain the following examples, it | — The in RCE x digital negative ae be obtained by using ‘The identity transformation is given in the Fig 43:1). 3 = 225-r(tmax=255) - ‘n the Fig. 43.1(a), the solid line is the transformation T. ~ Hence when ‘The horizontal 5 srt a (le a tpn on ¥ = 0,5=255 and when r=255, s = 0. sete optima aa Ingeneral s = (U-1)-r (431 ! ~ Wis cled an ee “ce ty en it docs | Mere Listhe number of grey levels.(256 in thiscase) 43.2 Con = - = = As seen the grey level | ~ "ataton (43.1) can be writen in terms of f(y) °°: — an modelo 10,1250 125 sd aly 256 gee, | 8A) = = poe Fle: £3.2(3) wil help us understand ey) = (L-1)- Ixy) (432) aa Processing techniques better. Here (x.y) the input mage and g (,y isthe ox! Teese te ‘image. portions da ed(431 se) y) 30 (432 utp MATLAB program for finding the digital negative 96% MATLAB code to calculate negative 94% clearall read (‘[Link]}; jouble (aa) 55; % for a 8-it image % figure(1) colormap(gray) imagesc(a) figure(2) eolormap(eray) imagesc(b) Fig. 4.3.3(b) : Digital negative 43.2 Contrast Stretching ~ Many times we obtain low contrast images due to poor ‘Mluminations or due to wrong setting of the lens aperture. The idea behind contrast stretching is to Increase the contrast of the images by making the dark portions darker and the bright portions brighter. 3 Image Enhancement in Spatial Do Fig. 434 shows the transformation used to achieve contrast stretching In the Fig 4:3.4, the dotted line indicates the identity transformation and the solid line is the contrast stretching transformation. ~ Asis evident from the Fig. 4.3.4, we make the dark grey levels darker by assigning a slope of less than one and make the bright grey levels brighter by assigning a slope greater than one, ~ One can assign different slopes depending on the input Image and the application, As was mentioned, image enhancement is a subjective technique and hence there is no one set of slope values that would yield the desired result 258 Identity w transformation Modified grey lavet i i 0 = > 255 Original grey level Fig. 4.3.4 : Original grey level r — The formulation of the contrast-stretching algorithm is given below. ee Oa es Here f (& y) is the input image and g (x, y) is the outp image ~ As mentioned earlier, image enhancement bei subjective phenomena, the value of @ will vary (rt Image to image and from person to person. 7 ‘objective Is to identify the region that he or she Interested in. An important thing to note is that thresholded image has the maximum contrast asi! 9.43.5): Orginal image only black and white grey values, ‘MATLAB program for thresholding %%6% Thresholding 49% 9% ear all de ‘imread(‘[Link]); sin row coll=size(f); pe nnaCtnter valu of Threshold) row forj=[Link]ol is 434 G = What th two part of grey vi X-ray or transforr ~ The tran: similar select at This can(xy) ax (436 1 outp being ary fron she s that as it hs W__image Processing if@j)sosatet) 159) 2(u)=255; else {ij)=0: end end end figure (1);% —-original image. imshow (p) : fiqure (2); % gray eve slicing without backgrow Imshow (uint8(2)) () Gray level slicing without background Fig. 439 6% Grey level slicing with backgoundl 9496 ear all de peimread(skulltif); ‘double(p}{row collsize(p); for'=t:1:row forjs i eee Feu sonente| i) 255: (u)-150) S else 2Gii)=POD: end end end ‘igure (1); % imshow(P) figure (2); tmshow (uint8(2)) - original image grey level slicing with back, (b) Grey level slicing with background Fig. 4.3.10 4.3.5 _ Bit Plane Slicing = In this technique, we find out the contribution made each bit to the final image. As mentioned earle image is defined as say a 256 x 256 x 8 image. In 256 256 is the number of pixels present in thei and 8 is the number of bits required to represen pixel. 8-bits simply means 28 or 256 grey levels Now each pixel will be represented by 8-its xample black is represented as 00000000 and wh represented as 11111111 and between them, 254 levels are accommodated. n bit plane slicing, wes ‘importance of each bit in the final image. This < done as follows, Consider the LSB value of ect P? and plot the image using only the LSBs. Continue doing this for each bt till we come tothe Note that we wil get 8 diferent images and al! Images will be binary. Ex. 4.31: Gv soln, to represer Hencowe' binary we 00 | om [| Binary mage So NATLAB dearall a 10 de asimread('wa adouble(a); rinput(whie SB); [row col}=si forx=E:L:row fory=1 c=dec? Aec(r) end end figure(1) ‘mshow(uint figure(2) ‘mshow(uintQn made earl, ge. Int the in resent B-bits. nd whi n, 254 -wesee this each pi tothe all Wings nce ; Ex. 4.3.1 : Given a 3 x3 image, plots bit planes. Soln. : — Since 7 \s the maximum grey level, we need only 3-bits to represent the grey levels. = Hence we will have 3-bit planes. Converting the image to binary we get, eor|ow|oo| |r{olo| |olsfo} jololo sooforr ow} fol r}o] jolatr| |rfolo anfirfoo} [ofr to} [+ folr] [sfrfo ‘Braryimage—“LSBplne Mile biplane SB plane %69 MATLAB code for bit extraction %% clearall de asimread('warnettif); a=double(a); r=input(‘which bit image do you want to see 1=MSB 8=1SB); [row col}-size(a); forx=1:l:row fory=[Link]ol ~— edecZbin{a(xy),8); % converts decimal to binary aec(r): w(sy)-double(d); 9696 since wis a char and cannot be plotted. ‘w(xy)==49 = %% since double of d will be either 49 or 48. w(xy)=255; else whxy)=0; aA 3 end figure(t) ae ‘Amshow(uint8(a)), figure(2) ‘ Amshow(uint8(w)) Image Enhancement in Spatial Domain Fig. 4.3.11 : Eight images, each representing contribution of a single bit — Observing the images we come to the conclusion that the higher order bits contain majority of the visually significant data, while the lower bits contain the suitable details in the image. = Bit plane slicing can hence be used in image ‘compression. We can transmit only the higher order bits and remove the lower order bits. Bit plane slicing is also used in steganography, Steganography : ~ Steganography is the art of hiding information, With the growth of networked multimedia systems, the need of secure data transfer increases. Steganography is a technique in which secret data is hidden in a carrier signal, in this case image. An intruder on the network sees the carrier image without realising that there is hidden information present init. ~ Bit plane manipulation is the simplest of the various steganography techniques available. We have seen that the bits representing the MSB's carry a lot of information while the bits representing the LSB's carry information that is visually insignificant. Consider an ‘example wherein we need to hide a text message in the image of a group. ~The image of the group is called a carrier image. We hide a text message (which is stored as an image) into the carrier image. What is done here is every LSB of the carrier image is replaced by the MSB of the secret data,= The final image that we obtain is called a stego image. It has the ae Image hidden into Ht but is visually ‘identical to the orignal mage. i tf we replace more than 2 LSBs with 2 MSB' we get the secret image superimposed on the original image. This is called a Watermark, Given below is the code for steganography and also for Enhancement in sp image Pre — This techn ‘Samir Adi “Transformves us ano? WD image Processing 4 {@) Carrier image Fig. 43.13 | Workers of the World, | Unite Lenin (b) Secret data — This technique of steganography is a very simplistic one. A lot of work has been done by my students Mustensir Lehri, Samir Adhia and Rahul Chedda from the computer department. You could send me an e-mail to view the paper “Transform based steganography’, published by them. (a) Stego image Workers of the World, Unite Lenin (b) Retrieved data Fig. 43.14 4,3.5(A) Applications of Bit Plane Slicing = Bit plane slicing is used in Steganography and Watermarking. Steganography is the art of hiding information, With the growth of networked multimedia systems, the need of secure data transfer increases. = Steganography is a technique in which secret is hidden in a carrier signal, in this image. An intruder on the network sees the carrier image without realising that there is hidden information presentin it. Dynamic Range Compression (Log Transformation) — At times, the dynamic range of the image exceeds the capability of the display device. What happens fs that some pixel values are so large that the other low value pixels get obscured, A simple day-to-day example of ‘such a phenomena is that during daytime, we cannot see the stars, 436 — The reason behind this is that the intensity of the sun is so large and that of the stars is so low that the eye cannot adjust to such a large dynamic range. — In image processing, a classic example of such large differences in grey levels is the Fourier spectrum (will ‘be discussed in detail in the frequency domain enhancement technique). Inthe Fourier spectrum only some of the values are very. large while most of the values are too small. The ‘dynamic range of pixels is of the order of 10%. Hence, whea we plot the Fourier spectrum, we see only small dots, which represent the large values. — Something needs to be done to be able to see the small values as well. This technique of compressing the ‘dynamic range is known as dynamic range compression. = We all know that the log operator is an excellent compressing function, Hence the dynamic range compression is achieved by using a log operator: C is the normalisation constant.4-10 mage Processing exbog(t+/") Fig. 43.15 = Dynamic range compression Ca fey) and els) as atay) = ex logtt+ifexy)D MATLAB program for dynamic range compression ‘4% Dynamic range compression %% learall de aazimread(saturntif); a=double(aa) [row col]=size(a); row for y=[Link] efxy)=a(xy)"(C-1)*Gs+y)); 96% Needed to center the transform end end azabs{ff2(0); [Link]( 140); 696% Plotting figure(1) colormap(eray) Jimagese(d) figure(2) colormap(gray) ‘imagesc(d_log) for: un be written in terms of Enbanc Ima (b) jig. 4.3.36 : Dynamic range compressio . Law Transformation for i=1:x« 43.7 Powel — the basic formula for power-law transforn sect itcan also be weitten a Pe stay) = exftay! end — Here cand y are positive constants. The tran nuimg2= is shown below for different values «i subplot(2 called the gamma correction factor. We imshow(i changing the value of gamma, we obtai tile(Orig transformation curves subplot(2 = Nonlineaites encountered during image « imshow(n title Imag printing and displaying can be corrected sing correction. Hence gamma correction is impor image needs to be displayed on the comp power law transformation can also be used t the dynamic range of an image. Given belo MATLAB code for power transformation mage has been normalized to 0- 255 range. Zr 438 Spatial Res Spatial a » pistes rn ants these a Intensity Re me = Intensity [row,col} ) resolution or [=size(img!) bs grey levels t Se ‘correction factor: resolution img=double(i img} the image a1 change in grssion portan d to in yelow 1. The W image Processing 4 W Image Enhancement in Spatial Domain for j=1:col snuimg(ij)=img(ij)*gamma end end numax=max(max(nuimg)) numin=min(min(nuimg)); 1n=255/(numax-numin} for i=t:row for j=1:col nuimgi (ij)=n*(nuimg())-numin), % Normalisation end end nulmg2=uint8(nuimg1); subplot(2,1.1) imshow(imgi) title(‘Original image’) subplot(2,1,2) imshow(nuimg2) title('Image after power transformation’) Fig. 43.18 4.3.8 Spatial and Intensity Resolution ‘Spatial Resolution : Spatial resolution is indicative of the number of samples that are present in the image. Hence Spatial resolution depends on the Sampling. Spatial resolution can be defined asis the smallest discernible detailin an image. Intensity Resolution : Intensity resolution which is also known as Gray level resolution or Tonal resolution is indicative of the number of ‘grey levels that are present in the image. Hence Intensity 4.4 Solved Examples on Point Processing Ex. 44.1: Obtain the digital negative of the following f gry lc 1 | 121 | 205 | at7 | 156 | 151 8 Bits Per Poel - BPP image | 130 | 127 | 157 | 117 | 125 198 | 142 197 | 242 240, Soln. : Its known that it sa 8-bit image. Hence the number of grey levels that this image can hold is 2°=256 «. L = 256. Hence the minimum grey level is hile the maximum grey level is 255 atxy) = (L-1)-ixy) = (256-1)-fixy) atxy) = 255-fIxy) ‘We gt the digital negative using the above equation 134 | 50 | 38 | 99 | 104 116 | 128 | 98 | 138 | 130 3 | 138 | 19 | 117 | 113 2 | 73 | 77 | s8 | 13 54 | 149 | 136 | 4 | 15 Ex. 442 : What would happen to the dynamic range of an Image it all the slopes in the contrast stretched algorithm (i,m, n) are tess than 1, (Answer using an example) Soln.: We know that contrast stretching generally Increases the dynamic range. Butin this case since , m, and n are less than 1. The dynamic range gets reduced. Let us take an example, Let = 0.2, m = 0.5, n = 02. Let the initial dynamic range of the original image be [0 - 10], letry= 4572-8 We draw this transformation resolution depends on the number of bits used to represent the image and can be defined as the smallest discernible change in gray level. br + -fmieraen nfr-1)+ 4