## Example Based Image Processing (2006)

Citations: | 3 - 2 self |

### BibTeX

@TECHREPORT{Gallagher06examplebased,

author = {Claire Gallagher},

title = {Example Based Image Processing},

institution = {},

year = {2006}

}

### OpenURL

### Abstract

ii This thesis is concerned with example based image processing. Example based image pro-cessing is a general term for any class of image processing operation where the manipulation and analysis of the image in question is guided by some set of example images. This thesis focuses on two applications, texture synthesis and image segmentation, in which example based image processing is proposed. Given an example texture, the goal of a successful texture synthesis algorithm is to generate new texture which is perceptually similar to the sample texture. One of the main challenges in this process is the modelling of the example texture. Previous work has shown that those algorithms based on implicit modelling are more successful than those based on the more rigid explicit models. Based on this observation a new texture synthesis algorithm is developed which combines the strength of the implicit modelling technique with wavelet based image analysis. The Dual-Tree Complex Wavelet Transform used in this work has associated with it good directional selectivity and shift invariance. Both of these properties make it well suited

### Citations

8919 | Maximum Likelihood from Incomplete Data via the EM-Alogrithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ...ting and refining the parameters associated with the Gaussian distributions, data can be partitioned based on their proximity to the nearest cluster space. The Expectation-Maximisation (EM) algorithm =-=[58]-=- is then used to determine the maximum likelihood parameter estimates for each Gaussian in the combined feature space. Probabilistic clustering algorithms can generate accurate segmentations of most i... |

7440 |
Probabilistie Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...nistic methods such as Iterative Conditional Modes (ICM) [22] were developed in order reduce the computationally load of MCMC methods. In recent years, Graph-Cuts [31] and Bayesian Belief Propagation =-=[155]-=- have been widely used in MAP-MRF estimation. 5.10.4.1 Markov Chain Monte Carlo Techniques For MAP-MRF estimation, the posterior distribution for the label field L to be estimated is given as, L ∼ p(L... |

6968 |
A mathematical theory of communication
- Shannon
- 1948
(Show Context)
Citation Context ... would then be to consider the joint probability of all pixels together. However, this is computationally intractable for images of realistic size. To overcome this Efros and Leung proposed a Shannon =-=[188]-=- inspired heuristic whereby the new texture image is grown outward in layers from an initial 3 × 3 seed taken randomly from the example image (in the case of region filling in or in-painting, the synt... |

4567 | A tutorial on hidden Markov models and selected applications in speech recognition
- Rabiner
- 1989
(Show Context)
Citation Context ...l may not be not be directly visible to the observer but variables influenced by the state are visible. The hidden Markov model (HMM) provides such a model. HMMs were first used in speech recognition =-=[165]-=- but have recently have been applied to many image processing problems. To illustrate how HMMs can be used in segmentation, consider the approach taken by Chen and Kundu [49] which aims to solve the p... |

4018 |
Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...eighbourhood at each resolution. Beginning at a coarse level and moving progressively toward finer resolutions, pixels are synthesised from the non-parametric MRF model using either the Gibbs sampler =-=[83]-=- or the more computationally attractive deterministic iterative conditional modes [21] algorithm. Both of these optimisation techniques will be discussed in more detail in Chapter 5. The algorithm des... |

3390 | A computational approach to edge detection
- Canny
- 1986
(Show Context)
Citation Context .... For example, the Sobel and Prewitt [85] operators are based on the first derivative and use a 3×3 convolution mask to approximate the absolute value of the gradient at each site in the image. Canny =-=[43]-=- developed a similar approach where the first derivate of a Gaussian was used to locate edges using the maximum gradient magnitude. To reduce the effect of noise on the segmentation, the raw image is ... |

2631 | A theory for multiresolution signal decomposition: The wavelet representation
- Mallat
- 1989
(Show Context)
Citation Context ...The suitability of the wavelet transform (WT) for use in image analysis has been well established. This is due to its ability to enable the localisation of functions in space, scaling and orientation =-=[141]-=-. In signal analysis, the WT can be interpreted as a signal decomposition onto a set of basis functions called wavelets. These wavelets can be described as small finite wavelike functions or ripples t... |

2286 |
A Wavelet Tour of Signal Processing
- Mallat
- 1999
(Show Context)
Citation Context ...time. This finite oscillating wavelike characteristic associated with wavelets makes them suitable for the analysis of transient, non-stationary or time varying phenomena associated with most signals =-=[38, 140]-=-. Figure 3.1 shows an example wave function (sinusoid) and a wavelet function (Morlet Wavelet). The wave function is smooth, periodic and infinite, whereas in contrast the wavelet function is finite a... |

2045 | Some methods for classification and analysis of multivariate observations
- MacQueen
- 1967
(Show Context)
Citation Context ...n an exclusive way, so that if a certain datum belongs to a definite cluster then it cannot be included in another cluster. One of the most popular hard clustering algorithms is the K−means algorithm =-=[139]-=-. K−means clustering is a simple and easy to implement technique for classifying a given data set into a certain number of clusters (assume K clusters) fixed a priori. The algorithm is developed based... |

1752 |
Vector Quantization and signal compression
- Gersho, Gray
- 1992
(Show Context)
Citation Context ...ssian pyramid individually. This is an even larger computational burden than that associated with Efros and Leung [69]. To reduce the computational expense, tree-structured vector quantisation (TSVQ) =-=[84]-=- is introduced as the searching algorithm [213]. This avoids the exhaustive nearest neighbour searching process by considering neighbourhoods as points in a multi-dimensional space and casting the nei... |

1411 |
Pattern Recognition with Fuzzy Objective Function Algorithms
- Bezdek
- 1981
(Show Context)
Citation Context ... each data point may belong to one or more clusters with different degrees of membership. In this case, data will be associated with an appropriate membership value. The fuzzy c−means (FCM) algorithm =-=[23, 67]-=- is the most widely used example of an overlapping clustering method. The algorithm is developed based on an iterative minimisation of the following objective function: U(g,f) = c� n� i=1 j=1 g m ij D... |

1400 | Color Indexing
- Swain, Ballard
- 1991
(Show Context)
Citation Context ...sed in more detail in section 5.6. Histograms are invariant to translation and rotation about the viewing axes. In general they change very little under varying scale and occlusion. Swain and Ballard =-=[201]-=- use the colour histogram for object indexing. Their algorithm is designed specifically for content baseds5.4. A Note on Features 96 retrieval where it is necessary to locate and compare similar objec... |

1316 |
An algorithm for vector quantizer design
- Linde, Buzo, et al.
- 1980
(Show Context)
Citation Context ...els. Figure 2.12 illustrates two example causal conditional neighbourhoods of the type used in [159]. A clustering based model is then trained on these vectors using a clustering process described in =-=[137]-=-. This clustering process is semi-supervised where the number of clusters must be specified beforehand and will be chosen based on the computational resources available together with the desired fidel... |

1239 |
Spatial Interaction and the Statistical Analysis of Lattice Systems (with Discussion
- Besag
- 1974
(Show Context)
Citation Context ...toward finer resolutions, pixels are synthesised from the non-parametric MRF model using either the Gibbs sampler [83] or the more computationally attractive deterministic iterative conditional modes =-=[21]-=- algorithm. Both of these optimisation techniques will be discussed in more detail in Chapter 5. The algorithm described in [152] is tested on both random and structured textures and the order of impr... |

1236 | Content-based image retrieval at the end of the early years
- Smeulders, Worring, et al.
(Show Context)
Citation Context ...and minor axes or/and the orientation of the major axis with the x−axis [44,105,195]. Other spatial methods sometimes used include moments, typically in the first (centre of gravity) or second (size) =-=[192]-=-. In general, using spatial features on its own will only work for the simplest of images. In addition, when considering implementation, the region must be identified before the shape can then be calc... |

1090 | The Laplacian pyramid as a compact image code
- Burt, Adelson
- 1983
(Show Context)
Citation Context ...on characterising a texture image can be captured in the first order statistics of a defined set of linear filter outputs. They model the texture in a multi-resolution domain using either a Laplacian =-=[39]-=- or Steerable pyramid structure [190]. An image pyramid is created by convolving and sub-sampling an image with a bank of linear filters. The defining characteristic of an image pyramid is that the ba... |

1015 | Video Google: A text retrieval approach to object matching in videos
- Sivic, Zisserman
- 2003
(Show Context)
Citation Context ...means of judgment on the outcome. As a compromise much of the recent segmentation research has been focused on semi-automated segmentation where some clue as to the image content is given as an input =-=[90,131,147,175,179,191]-=-. Semi-automated segmentation can be considered as the interleaving of object recognition and segmentation and the task is now that of: given an example object, does this object exist in the image and... |

987 | On the statistical analysis of dirty pictures
- Besag
- 1986
(Show Context)
Citation Context ...main class of stochastic techniques used are Markov Chain Monte Carlo (MCMC) methods, of which simulated annealing (SA) is one example. Deterministic methods such as Iterative Conditional Modes (ICM) =-=[22]-=- were developed in order reduce the computationally load of MCMC methods. In recent years, Graph-Cuts [31] and Bayesian Belief Propagation [155] have been widely used in MAP-MRF estimation. 5.10.4.1 M... |

948 | Theory of edge detection
- Marr, Hildreth
- 1980
(Show Context)
Citation Context ...ed. A map of intensity and direction gradients are thus produced. The algorithm then uses thresholding to determine which points are edges by following edge lines through the image. Marr and Hildreth =-=[143,144]-=- found that improved results can be achieved using the second derivative of the image. These Laplacian operators locate zeros based on the rate of zero crossing in the second derivative. By definition... |

940 |
Theory of communication
- Gabor
- 1946
(Show Context)
Citation Context ... frequency localisation necessary for most naturally occurring (non-stationary) signals. To address the limitations of the standard FT, the Short Time Fourier Transform (STFT) was introduced by Gabor =-=[76]-=-. The STFT represents a sort of compromise between the time and frequency based views of a signal. Essentially, the STFT moves a fixed window over the length of the signal. Inside this window, the sig... |

900 | An introduction to hidden Markov models
- Rabiner, Juang
- 1986
(Show Context)
Citation Context ...ypes of MRF models have been used but three types have remained dominant. These are the Gaussian Markov Random Field (GMRF) [47], the Potts model [162] and more recently the Hidden Markov Model (HMM) =-=[164]-=-. The GMRF was discussed earlier as the prior on the texture model and so will not be discussed in this section. The Potts model can be considered as a classical Markov model, which aims to model only... |

817 | Texture synthesis by nonparametric sampling
- Efros, Leung
- 1999
(Show Context)
Citation Context ...ametric energy based modelling. Using the input parameters of (i) observed intensity image to be segmented I, (ii) training example image Ie and (iii) example segmentation of image (ii) Le. and Leung =-=[69]-=- demonstrated the strength of this heuristic technique in their non-parametric texture synthesis algorithm which was discussed in Chapters 2 and 4. To synthesise a new texture image, Efros and Leung p... |

742 | Statistical pattern recognition: a review
- Jain
- 2001
(Show Context)
Citation Context ...l focus on those techniques that have had the biggest influence on thes5.6. Thresholding 99 work carried out in this thesis. A more detailed review of existing segmentation algorithms can be found in =-=[65,97,110,150,153,185]-=-. Note that because of the wide variability in image types, it has been found that no one approach is suitable for all image types. As a consequence of this general observation, a successful segmentat... |

703 | GrabCut” - interactive foreground extraction using iterated graph cuts
- Rother, Kolmogorov, et al.
- 2004
(Show Context)
Citation Context ...means of judgment on the outcome. As a compromise much of the recent segmentation research has been focused on semi-automated segmentation where some clue as to the image content is given as an input =-=[90,131,147,175,179,191]-=-. Semi-automated segmentation can be considered as the interleaving of object recognition and segmentation and the task is now that of: given an example object, does this object exist in the image and... |

697 | Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images
- Boykov, Jolly
- 2001
(Show Context)
Citation Context ...nnealing (SA) is one example. Deterministic methods such as Iterative Conditional Modes (ICM) [22] were developed in order reduce the computationally load of MCMC methods. In recent years, Graph-Cuts =-=[31]-=- and Bayesian Belief Propagation [155] have been widely used in MAP-MRF estimation. 5.10.4.1 Markov Chain Monte Carlo Techniques For MAP-MRF estimation, the posterior distribution for the label field ... |

697 |
The structure of images
- Koenderink
- 1984
(Show Context)
Citation Context ...ge allows both class and spatial information to be distributed across various scales. This results in the provision of a simple hierarchical framework for detailed image analysis of image information =-=[123]-=-. Typically, a multi-scale representation will reduce the spatial resolution of the image by a factor of two at each scale resulting in a pyramid of successively smaller but similar images. A simplifi... |

668 | VisualSEEk: A fully automated content-based image query system
- Smith, Chang
- 1996
(Show Context)
Citation Context ...ellipse to describe region shape. The feature descriptors used could then take the form of the ratio of the length of the major and minor axes or/and the orientation of the major axis with the x−axis =-=[44,105,195]-=-. Other spatial methods sometimes used include moments, typically in the first (centre of gravity) or second (size) [192]. In general, using spatial features on its own will only work for the simplest... |

633 | An algorithm for finding best matches in logarithmic expected time
- Friedman, Bentley, et al.
- 1977
(Show Context)
Citation Context ...ndency of [69], however it does not remove it entirely. The computational burden associated with the exhaustive nearest neighbour searching inherent in [69] is also removed in [99]. Rather, a kd-tree =-=[75]-=- struc2.5. Non-Parametric Approaches 29 ture is introduced in order to reduce the search time. As with most nearest neighbour searching approximations, the reduction in search time comes at the expen... |

595 |
Statistical and Structural Approaches to Texture
- Haralick
- 1979
(Show Context)
Citation Context ...ture by which the identification of regions can take place. There have been many different approaches proposed to extract textural information from images. Review papers on texture feature extraction =-=[95, 167, 210]-=- conclude that because of the wide variability in texture behavior, no one method may be considered suitable for all texture types [209]. Traditionally, the first approaches to texture analysis were s... |

566 |
The wavelet transform, time-frequency localization and signal analysis
- Daubechies
- 1990
(Show Context)
Citation Context ...ave shape is independent of scale and is known as the mother wavelet. The WT has found many applications in image processing due to its ability to localise functions in space, scaling and orientation =-=[57,60,141,170]-=-. Some example applications of the WT are in image compression [8,9,107,205], image de-noising [62,91,176], digital watermarking of images [138], object segmentation [51,187,199] and texture analysis ... |

563 |
Theory of Probability
- Jeffreys
- 1948
(Show Context)
Citation Context ...stimates. Based on this observation it is found that model simplicity is the key to maximising the degree of consistency between parameter estimates computed from independent realisations of the data =-=[112,177]-=-. The computation of the model selection evidence is difficult as it involves integrating the product of the likelihood and the prior probability density functions over all the possible parameters in ... |

543 | Image quilting for texture synthesis and transfer
- Efros, Freeman
- 2001
(Show Context)
Citation Context ... of horizontal and vertical edges. The manner in which each diamond block is chosen and merged with others is derived from the image quilting texture synthesis algorithm proposed by Efros and Freeman =-=[68]-=-. Using Wang tiles for texture synthesis works well on random textures that contain little structure. However, for more structured textures, a larger set of tiles would have to be used thus reducing t... |

534 | Distortion invariant object recognition in the dynamic link architecture
- Lades, Vorbruggen, et al.
- 1993
(Show Context)
Citation Context ...tween neighbourhoods. Other possible suggestions including comparing the energy between neighbourhoods or creating histograms and comparing neighbourhood histograms. The suitability of wavelet “jets” =-=[133, 218]-=- as a means of image representation will also be investigated. 6.8.4 Comparison with Mignotte Figure 6.20 shows two segmentations of the observed image in (i). Segmentation (iii) is the MAP estimate o... |

499 | Wavelets and subband coding
- Vetterli, Kovacevic
- 1995
(Show Context)
Citation Context ...t theory. Therefore, topics such as the mathematical foundations of the transform, and wavelet filter design will not be discussed here. However, detailed descriptions of these topics can be found in =-=[119,122,129,140,170,212]-=-. 3.1 The Discrete Wavelet Transform The basic building blocks of the one dimensional (1D) Discrete Wavelet Transform (DWT) are shown in Figure 3.3. During the analysis stage of the DWT, the input sig... |

494 | Unsupervised texture segmentation using gabor filters
- Jain, Farrokhnia
- 1991
(Show Context)
Citation Context ...of the area of the frequency domain which it analyses. In texture applications, banks of these filters are chosen to analyse the particular spatial frequencies and orientations present in the texture =-=[30, 111]-=-. Although Gabor filters have shown considerable success in texture analysis, their widespread use has been hampered by two large inherent limitations associated with their use. Firstly, the filter pa... |

477 |
Chistoph von der Malsburg. Face recognition by elastic bunch graph matching. 96-08, Institut f"ur Neuroinformatik
- Wiskott, Fellous, et al.
- 1996
(Show Context)
Citation Context ...tween neighbourhoods. Other possible suggestions including comparing the energy between neighbourhoods or creating histograms and comparing neighbourhood histograms. The suitability of wavelet “jets” =-=[133, 218]-=- as a means of image representation will also be investigated. 6.8.4 Comparison with Mignotte Figure 6.20 shows two segmentations of the observed image in (i). Segmentation (iii) is the MAP estimate o... |

462 | Shiftable Multi-scale Transforms - Simoncelli, Freeman, et al. - 1992 |

450 | Fast texture synthesis using tree-structured vector quantization
- WEI, LEVOY
(Show Context)
Citation Context ...e same powerful modelling framework, then the resulting algorithm would provide a complete solution to texture synthesis problem.s2.5. Non-Parametric Approaches 26 Fast Synthesis Using Image Pyramids =-=[214]-=-, (2000) Around the same time that Efros and Leung published their algorithm, Wei and Levoy [214] published a similar algorithm which, as the authors state 3 , was not derived from [69] but rather ext... |

422 |
Multichannel texture analysis using localized spatial filters
- Bovik, Clark, et al.
- 1990
(Show Context)
Citation Context ...of the area of the frequency domain which it analyses. In texture applications, banks of these filters are chosen to analyse the particular spatial frequencies and orientations present in the texture =-=[30, 111]-=-. Although Gabor filters have shown considerable success in texture analysis, their widespread use has been hampered by two large inherent limitations associated with their use. Firstly, the filter pa... |

383 |
Introduction to Wavelets and Wavelet Transforms
- Burrus, Gopinath, et al.
- 1998
(Show Context)
Citation Context ...time. This finite oscillating wavelike characteristic associated with wavelets makes them suitable for the analysis of transient, non-stationary or time varying phenomena associated with most signals =-=[38, 140]-=-. Figure 3.1 shows an example wave function (sinusoid) and a wavelet function (Morlet Wavelet). The wave function is smooth, periodic and infinite, whereas in contrast the wavelet function is finite a... |

372 | Image analogies
- Hertzmann
- 2001
(Show Context)
Citation Context ...se of the L− shape causal neighbourhood used in the synthesis process. (i) (ii) (iii) Figure 2.21: Synthesised textures using the Ashikhmin [14] algorithm. Combining Coherent and Exhaustive Searching =-=[102]-=-, (2001) To avoid the inclusion of artificial horizontal edges in the synthesised result, Hertzmann et al. [102] combines the exhaustive nearest searching proposed by Wei and Levoy with the coherents2... |

369 | Graphcut textures: Image and video synthesis using graph cuts
- Kwatra
- 2003
(Show Context)
Citation Context ...resents a pixel at location x in Is, where x ∈ X. Existing texture synthesis algorithms that generate an output texture image from a given sample texture can be roughly categorised into three classes =-=[132]-=-: parametric, non-parametric and patch based. Some of the approaches that are especially pertinent to this work have been implemented and synthesised results are shown. In order to appreciate the subt... |

366 |
Fundamentals of digital image processing
- Jain
- 1989
(Show Context)
Citation Context ...up the synthesis process.s4.3. DT-CWT TexSyn 73 2. Luminance then Chrominance Synthesis: The sensitivity of the human eye to luminance or brightness fluctuations in an image has been well established =-=[109]-=- and exploited for years through black and white images and video representation. To isolate the luminance components of an image, the image is moved from RGB colour space to a colour space which has ... |

357 | Blobworld: Image segmentation using expectationmaximization and its application to image querying
- Carson, Belongle, et al.
- 2002
(Show Context)
Citation Context ...cognition. By automatically detecting and separating objects in images, large databases can be searched without the need for manual intervention. This is an important issue in content-based retrieval =-=[10,44,186]-=- and in image compression [40,168]. 86s5.2. Segmentation: The User’s Interpretation of a Scene 87 (i) Observed image (intensities) (ii) Segmented image (labels) Figure 5.1: Segmentation is the process... |

350 |
A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters
- Dunn
- 1973
(Show Context)
Citation Context ... each data point may belong to one or more clusters with different degrees of membership. In this case, data will be associated with an appropriate membership value. The fuzzy c−means (FCM) algorithm =-=[23, 67]-=- is the most widely used example of an overlapping clustering method. The algorithm is developed based on an iterative minimisation of the following objective function: U(g,f) = c� n� i=1 j=1 g m ij D... |

344 |
Visual Information Retrieval
- Bimbo
(Show Context)
Citation Context ...le, the HSV colour space format has been widely used in many image analysis applications [193]. An alternative colour space that is popular in content based image retrieval systems is the L ∗ a ∗ b ∗ =-=[25]-=- colour space. Similar to HSV , the transformation from RGB to L ∗ a ∗ b is a nonlinear and the L ∗ a ∗ b colour space has been explicitly designed to achieve perceptual uniformity. The luminance chan... |

344 |
Exact maximum a posteriori estimation for binary images
- Greig, Porteous, et al.
- 1989
(Show Context)
Citation Context ...ticular application to the binary (two label) segmentation problem, techniques based on graph-cuts have been introduced into the segmentation framework to minimise the energy function given by (5.18) =-=[31,88,128]-=-. The idea behind the graph-cut techniques5.10. The Bayesian Segmentation Framework 121 is to construct a specialised graph for the energy function to be minimised such that the minimum cut on the gra... |

339 |
Poisson image editing
- Pérez, Gangnet, et al.
- 2003
(Show Context)
Citation Context ...ions. Normally areas to be filled in with new texture are small and the greatest concern is to interpolate smoothly between the boundaries and new synthesised region. Descriptions of these algorithms =-=[20,89,124,156,211]-=- will not be discussed here as they are generally not suitable for synthesising large regions. Figure 2.6 (ii) shows the other type of texture synthesis problem which is the focus of this work. Let Ie... |

337 | Wavelet-Based Statistical Signal Processing Using Hidden Markov Models
- Crouse, Nowak, et al.
- 1998
(Show Context)
Citation Context ...of the basic MRF model and can essentially be classified as a doubly stochastic random field in which there exists a hidden parameter whose value in not directly observable to the user. Crouse et al. =-=[56]-=- proposed the use of the HMM in order to characterise the wavelet dependencies among wavelet coefficients at different levels of the wavelet tree. In the Fan and Xia algorithm the HMT-3S model is base... |

325 | Blobworld: A system for region-based image indexing and retrieval
- Carson, Thomas, et al.
- 1999
(Show Context)
Citation Context ... mixture of these distributions. The parametric approach is generally based on the assumption that the underlying data density can be modelled using some definable process. For example, Carson et al. =-=[45]-=- use a mixture of k Gaussian densities modelled in an 8D feature space. The feature space is composed of a mixture of components from colour, texture and positional features. By estimating and refinin... |