ADVANCE IMAGE PROCESSING TECHNIQUES

Chapter 1 - Joint Edge-Preserving and Moiré-Suppressing Image Resampling Using the Discrete Gabor Transform
D. Van de Ville, W. Philips, and I. Lemahieu

Many printing techniques today use classical halftoning to create the illusion of contones on bi-level printing devices. Halftone dots are placed on a regular lattice. Resampling the original image to this lattice can introduce dreaded moiré artifacts in the halftoned images due to aliasing if the original image contains high frequency patterns. Using linear resampling filters with low-pass characteristics can prevent aliasing but also unacceptably blurs the image. Non-linear resampling schemes allow combining both properties to obtain better image quality.

Part 1 of this chapter introduces sampling theory for general non-separable periodic lattices. In particular, the case of gravure printing is considered in more detail. Gravure printing uses a low-resolution hexagonal lattice and has only a limited degree of freedom in constructing the halftone dots due to the physical constraints of the engraving mechanism. The nature of the moiré patterns is analyzed in detail. We show that linear interpolation functions must make a compromise between edge-preservation and anti-aliasing.

Part 2 explains how a local estimate of the risk of aliasing can be computed from the original images once the target lattice is known. This estimate is calculated using a space-frequency representation by computing the discrete Gabor transform coefficients. An important criterion for the risk estimation algorithm that is investigated, is the response on a pattern with increasing frequency. Finally, the risk can be used as a linear weighting factor between two appropriate linear resampling filters, i.e., an edge-preserving filter and a moiré-suppressing filter.

In the course of this chapter several results demonstrate the properties of the different approaches. As this book is not produced by gravure printing, we simulate the gravure printing to demonstrate the experimental results.

Chapter 2 - Scanned Multiresolution Imaging Using Binary Hadamard Mask
Don Bone, and Dan C. Popescu

One of the most significant impediments to the wider adoption of electronic imaging is the cost of photographic resolution sensors. Various schemes have been developed in the past to use masking techniques on the elements of a low cost, low resolution sensor which can then be used to produce high resolution images from multiple displaced exposures on the sensor. These devices can suffer from poor light response, difficulty in being adapted to a 2-dimensional sensor, and non-uniform noise response. This paper presents an approach based on a new class of Hadamard masks which we call Uniform Noise Hadamard Masks. In addition to the improved response over simple sampling approaches such as that used in the Kontron^TM cameras, they exhibit multiresolution capabilities and uniform noise response.

Chapter 3 - The Application of Markov Random Field Models to Wavelet-Based Image Denoising
Alexandra Pizurica, Wilfried Philips, Igance Lemahieu and Marc Acheroy

In this chapter, we address the use of Markov Random Field (MRF) prior models in wavelet based image denoising. Two different approaches are considered: the maximum a posteriori (MAP) estimation of the wavelet coefficients using MRF priors, and the MAP estimation of significant edges in detail images. The second approach can be incorporated into many different wavelet based denoising schemes; in particular, we discuss its use in wavelet shrinkage techniques. Practical results demonstrate that this approach is very efficient for suppressing different kinds of natural noise, which makes it attractive for different applications.

Chapter 4 - Characteristic Interaction Structures in Gibbs Texture Modeling
Georgy Gimel'farb

Under the assumption that spatially homogeneous image textures are modeled by Gibbs random fields with multiple pair-wise pixel interactions, parallel and sequential schemes of learning a characteristic interaction structure are compared. The learning can be based on a comparison of relative interaction energies or other integral characteristics of each interaction, for instance, chi-square distances between certain grey level difference or co-occurrence histograms collected over a given training sample. Parallel thresholding selects a basic structure of stronger, that is, more energetic or distant, interactions that is sufficient to model a specific class of stochastic textures introduced in [3, 4]. In many cases the basic structure is larger than is required because it contains not only the strong primary interactions but also the sufficiently strong secondary interactions obtained by a statistical interplay of the primary ones. Empirical sequential learning proposed by Zalesny [6, 7] tends to exclude the secondary interactions so that the basic structure can be reduced in size and complemented with a fine structure that describes minor but visually important repetitive details of a texture. The empirical sequential scheme that involves a great body of computations can be approximated by less complex analytical and combined analytical-empirical sequential schemes. Experiments show that the sequential learning results in more precise Gibbs models of non-stochastic regular textures but does not improve (and sometimes may even deteriorate) the basic structure of stochastic textures.

Chapter 5 - Multifractal Multipermuted Multinomial Measures for Texture Characterization and Segmentation
Lui Kam and Jacques Blanc-Talon

This chapter deals with the problem of texture approximation by means of a multifractal model allowing the unsupervised segmentation of real world images. By extending multinomial measures, a new class of self-similar multifractal measures is developed for this purpose. Two multifractal features have been shown to be suitable for texture discrimination and classification. Their use within a supervised segmentation framework provides us with satisfactory results.

Here, we complete the survey on these features by showing their rotation invariant property and their scaling behaviour. Both properties are particularly important for analyzing aerial and satellite images because the geographical elements can appear in different orientations and scales. Then, an automatic clustering algorithm based on a watershed technique is used for the segmentation of real world images. Experimental results obtained in different vision applications reveal the efficiency as well as the limits of our approach.

PERFORMANCE EVALUATION IN COMPUTER VISION

Chapter 6 - Performance Characterization in Computer Vision: The Role of Statistics in Testing and Design
Patrick Courtney and Neil A. Thacker

We consider the relationship between the performance characteristics of vision algorithms and algorithm design. In the first part we discuss the issues involved in testing. A description of good practice is given covering test objectives, test data, test metrics and the test protocol. In the second part we discuss aspects of good algorithmic design including understanding of the statistical properties of data and common algorithmic operations, and suggest how some common problems may be overcome.

Chapter 7 - Quantitative Assessment of Image Filtering: Comparison of Objective Metrics
Didier Coquin and Phillippe Bolon

Image processing consists of handling both intensity-energy information (grey-level) and geometric information (shapes). Hence, assessment operators have to take both effects into account. In this paper, five objective quality metrics are investigated and their behavior studied in terms of sensitivity to noise and to shape variations.

Chapter 8 - Progressive Image Coding for Improved Perceptual Quality and Recognition at Low Bit Rates
Dirck Schilling and Pamela Cosman

An important consideration facing designers of progressive image coders is how to assign priority in the transmitted bit stream, such that the "most important" information is sent first. PSNR (peak signal-to-noise ratio) is often used for this purpose, assigning the greatest priority to those bits which yield the greatest reduction in mean squared error. PSNR often fails, however, to adequately reflect the perceptual quality of compressed images. For several types of algorithms, including those with spatially scalable decoders, PSNR might not even be computable, and other methods of evaluation must be used.

In this paper, we present two progressive coding algorithms designed to improve the visual appearance and recognizability of compressed images at very low bit rates. The first coder is a variation of SPIHT, modified to make it spatially scalable without any loss in performance or in progressivity. We present experimental results comparing this multiscale SPIHT (MSPIHT) against SPIHT in terms of the bit rates at which viewers recognize objects in the reconstructed images. We show that viewers are able to recognize reduced-scale images, such as those compressed by MSPIHT, substantially earlier than images compressed by SPIHT. The second algorithm employs edge enhancement to improve the perceptual quality of compressed images. Important edges in the original image are transmitted together with a traditional wavelet coder bit stream. The decoder combines the two complementary information sources in a manner which, for certain image classes, can yield highly recognizable images at very low bit rates.

Chapter 9 - On the Masking Effects in a Perceptually Based Image Quality Metric
Abdelhakim Saadane, Nachida Bekkat, and Dominique Barba

The necessity of having at disposal an algorithm to control and assess the quality of digital images and videos is well-known from the community of broadcasters and scientists. This paper presents a new objective quality evaluation scheme. Human Visual System (HVS) properties are considered in the design of each part of the scheme. Compared to the well approved existing ones, the novelty of this paper concerns the masking effects modeling. Based on the measure of maximum quantization steps without visual impairments rather than the differential visibility thresholds, the approach described here considers both intra-channel and inter-channel masking. Experiments have been conducted on complex images (noise and texture) instead of simple sinusoidal stimuli. A model which describes the observed dependencies is derived and then included in an objective quality metric. Its different parameters have been optimized in order to maximize the usual correlation as well as the ranking correlation between the objective measures given by the method and the subjective scores (MOS) obtained from psychophysical experiments.

INTELLIGENT VISION SYSTEMS

Chapter 10 - Advanced Computer Vision and Graphics Collaboration Techniques for Image-Based Rendering
Samuel Boivin and Andre Gagalowicz

The idea of using real images to generate photorealistic computer graphics (scenes) has led to the development of Image-based modeling and rendering. These techniques are very similar to those developed some years ago within the framework of analysis/synthesis collaboration. In this paper, we present a new approach to reconstruct the 3D geometry and photometry of a scene based upon two distinct processes. A vision process uses two digital images of a scene captured with a camera to reconstruct its full geometry. A computer graphics process uses a single image to recover the photometry of the surfaces (i.e. the surface reflectances) and the radiance-to-pixel function by minimizing an error function. The generated images are then used as a feedback to modify the surface reflectances. Our aim is to find the simplest reflectance model allowing to faithfully reconstruct the original image of a scene, keeping in mind that the related photometric analysis is highly dependent on the complexity of the searched model.

Our approach generates photorealistic images using a rapid global illumination algorithm including the computation of a specular component. This algorithm is driven by the mean square error between the real image and the synthetic one, and minimizes it with respect to the parameters of the photometric model and of the radiance-to-pixel conversion functions. Several applications of this method are presented, such as augmented reality.

Chapter 11 - Adaptive Dynamic Scene Analysis
John F. Hadddon and Jimmy F. Boyce

Many future image analysis systems will involve a complex interaction between the algorithm, the algorithm's objective and the imaging environment. Currently, small changes in the imaging environment frequently have a disproportionate impact on the efficacy of the algorithm. The future development of real-world image analysis techniques will need to include techniques which enable the algorithms to adapt dynamically to changes in prevailing conditions, coupled with the capability to adapt to changing goals. This chapter presents current research at the Defence Evaluation and Research Agency into the incorporation of 'intelligence' into image processing techniques for scene analysis applications. A variety of techniques are being investigated which will enable the components of the image analysis systems to assess their own performance, to provide feedback to other components and to adjust process parameters to optimize performance for the prevailing imaging conditions, data and operational objectives. These concepts are leading to the development of dynamic and adaptive techniques with new levels of robustness.

Chapter 12 - Assessment of Image Processing Algorithms as the Keystone of Autonomous Robot Control Architectures
André Dalgalarrondo and Dominique Luzeaux

In this paper, we are interested in the design and the experiment of a control architecture for an autonomous outdoor mobile robot which mainly uses vision. We focus on the design of a mechanism that permits the dynamic selection and firing of perception processes, while managing the computational resources and allowing a fair reactivity. We show how such a construction relies deeply on the assessment of the image processing algorithms used by the various processes involved within the robot control architecture. We conclude with implementation issues and an experiment with a real-world robot in outdoor environments.

Chapter 13 - Parameter Optimization for Visual Obstacle Detection Using a Derandomized Evolution Strategy
Thomas Bergener, Carsten Bruckhoff and Christian Igel

The autonomous mobile robot ARNOLD uses information from a stereo camera system for navigating in an unknown and dynamically changing environment. A method called Inverse Perspective Mapping (IPM) is used for visual obstacle detection. The performance of this algorithm depends on the quality of the internal camera model. In this paper we employ an Evolutionary Algorithm (EA) to improve the parameters of this model. We use a derandomized evolution strategy called (μ/µ_I, λ)-CMA, which adapts the complete covariance matrix of the mutation distribution. After descriptions of the IPM and the CMA, we show that the proposed optimization method leads to better parameter settings than adjustment by an expert.

Chapter 14 - Obstacle Tracking in Registered Range and Reflectance Image Sequences Using Multivalued Template Matching
Karin Sabottka, Cristina Cris, Peter Zuber and Horst Bunke

The number and severity of traffic accidents can be reduced not only by safety systems such as seat belts and airbags, but also by advanced driver support systems. For this reason, research and the automotive industry have focused their attention more and more on such systems recently. Although our world is three-dimensional, most approaches in vision-based driver assistance have utilized grey level or color image sequences. However, the spatial arrangement of objects within a scene is often more relevant than the reflected brightness information. For this reason there has recently been an increasing interest in range sensors for collision avoidance systems.

For automotive applications sensor fusion offers several advantages. Different kinds of data or algorithms, respectively, can be used to overcome the limitations inherent in the use of single kind of data or algorithm and increase the overall robustness of a system. Intrasensor fusion will be in focus of this book chapter. As an example for intrasensor fusion, the problem of how to fuse range and reflectance data for obstacle tracking is studied. A novel approach to obstacle tracking in registered range and reflectance image sequences using a multivalued template matching scheme is proposed. The multivalued template of an object consists of two parts. The shape-based part contains range distances relative to a reference point on the object surface. Accordingly it describes the 3D shape of the tracked object. The reflectance-based part of the template contains information about the different reflective properties of the tracked object's surface materials. Range and reflectance data are fused to increase the reliability of obstacle tracking. For matching, the proposed approach takes into account particular problems with range data, normally not present in brightness imagery. The range and reflectance images are spatially under sampled and due to poor target reflectivity they include undefined data points. For these reasons tracking of objects is difficult and needs innovative techniques in the field of image processing. Experimental results are shown for two test sequences and are compared to the cases where only one of the information sources is used.