Image Generation Model Research Articles

A model for image generation and symbol recognition through the deformation of lineal shapes

We describe a general framework for the recognition of distorted images of lineal shapes, which relies on three items: a model to represent lineal shapes and their deformations, a model for the generation of distorted binary images and the combination of both models in a common probabilistic framework, where the generation of deformations is related to an internal energy, and the generation of binary images to an external energy. Then, recognition consists in the minimization of a global energy function, performed by using the EM algorithm. This general framework has been applied to the recognition of hand-drawn lineal symbols in graphic documents.

Computer vision applied to super resolution

Super-resolution (SR) restoration aims to solve the following problem: given a set of observed images, estimate an image at a higher resolution than is present in any of the individual images. Where the application of this technique differs in computer vision from other fields is in the variety and severity of the registration transformation between the images. In particular this transformation is generally unknown, and a significant component of solving the SR problem in computer vision is the estimation of the transformation. The transformation may have a simple parametric form, or it may be scene dependent and have to be estimated for every point. In either case the transformation is estimated directly and automatically from the images. We describe the two key components that are necessary for successful SR restoration: the accurate alignment or registration of the LR images and the formulation of an SR estimator that uses a generative image model together with a prior model of the super-resolved image itself. As with many other problems in computer vision, these different aspects are tackled in a robust, statistical framework.

Recognizing handwritten digits using hierarchical products of experts

The product of experts learning procedure can discover a set of stochastic binary features that constitute a nonlinear generative model of handwritten images of digits. The quality of generative models learned in this way can be assessed by learning a separate model for each class of digit and then comparing the unnormalized probabilities of test images under the 10 different class-specific models. To improve discriminative performance, a hierarchy of separate models can be learned, for each digit class. Each model in the hierarchy learns a layer of binary feature detectors that model the probability distribution of vectors of activity of feature detectors in the layer below. The models in the hierarchy are trained sequentially and each model uses a layer of binary feature detectors to learn a generative model of the patterns of feature activities in the preceding layer. After training, each layer of feature detectors produces a separate, unnormalized log probability score. With three layers of feature detectors for each of the 10 digit classes, a test image produces 30 scores which can be used as inputs to a supervised, logistic classification network that is trained on separate data.

Scene dependence of the non-Gaussian scaling properties of natural images

We report results on the scaling properties of changes in contrast of natural images in different visual environments. This study confirms the existence, in a vast class of images, of a multiplicative process relating the variations in contrast seen at two different scales, as was found in \cite{NIPS97,TuMaPaNa98}. But it also shows that the scaling exponents are not universal: Even if most images follow the same type of statistics, they do it with different values of the distribution parameters. Motivated by these results, we also present the analysis of a generative model of images that reproduces those properties and that has the correct power spectrum. Possible implications for visual processing are also discussed.

Scene dependence of the non-Gaussian scaling properties of natural images

We report results on the scaling properties of changes in contrast of natural images in different visual environments. This study confirms the existence, in a vast class of images, of a multiplicative process relating the variations in contrast seen at two different scales, as was found in Turiel et al (Turiel A, Mato G, Parga N and Nadal J-P 1998 Self-Similarity Properties of Natural Images: Proc. NIPS'97 (Cambridge, MA: MIT Press), Turiel A, Mato G, Parga N and Nadal J-P 1998 Phys. Rev. Lett. 80 1098-101). But it also shows that the scaling exponents are not universal: even if most images follow the same type of statistics, they do it with different values of the distribution parameters. Motivated by these results, we also present the analysis of a generative model of images that reproduces those properties and that has the correct power spectrum. Possible implications for visual processing are also discussed.

The electric image in weakly electric fish: physical images of resistive objects in Gnathonemus petersii.

The present study describes a measurement-based model of electric image generation in the weakly electric mormyrid fish Gnathonemus petersii. Measurements of skin impedance, internal resistivity and fish body dimensions have been used to generate an electrical-equivalent model of the fish and to calculate electrical images and equivalent dipole sources for elementary resistive objects. These calculations allow us to understand how exafferent and reafferent signals are sensed by electroreceptors. An object's electric image consists of the modulation of the transcutaneous voltage profile generated by the fish's own discharge. The results suggest a set of rules for electrolocation: (1) the side of the fish where modulation is larger indicates the side on which the object is situated; (2) the object's position in the electroreceptive field is indicated by the point of maximum modulation of the transcutaneous voltage; (3) the degree of focus of the image indicates the distance to the object. In addition, center-surround opposition originating at pre-receptor level is proposed. Both experimental measurements and modeling indicate that fish skin impedance is relatively low (400-11 000 <IMG src="/images/symbols/capomega.gif" WIDTH="13" HEIGHT="13" ALIGN="BOTTOM" NATURALSIZEFLAG="3">cm<SUP>2</SUP>) and mainly resistive. This low skin impedance appears to enhance the local electric organ discharge modulation, the center-surround effect, the signal-to-noise ratio for electrolocation and the active space for electrocommunication.

Mental images: Specific or general, personal or impersonal?

In three experiments the relation between visual images classified in terms of the four categories of personal, impersonal, specific, and general images was explored. The results showed that subjects were able to generate all types of images, with personal images being the easiest to recall. Personal images, together with general images, were the image types being easiest to integrate with one another. Overall, the observations suggested the existence of two imagery dimensions: specific‐general, and personal‐impersonal. Memory was found to be best for personal images. Personal images turned out to be easier to integrate in memory than specific images. The findings were discussed in terms of different image generation models.

A Model for Segmentation and Analysis of Noisy Images

Abstract This article proposes a statistical model for image generation that provides automatic segmentation of images into intensity-differentiated regions and facilitates the quantitative assessment of uncertainty associated with identified image features. The model is specified hierarchically within the Bayesian paradigm. At the lowest level in the hierarchy, a Gibbs distribution is used to specify a probability distribution on the space of all possible partitions of the discretized image scene. An important feature of this distribution is that the number of partitioning elements, or image regions, is not specified a priori. At higher levels in the hierarchical specification, random variables representing emission intensities are associated with regions and pixels. Observations are assumed to be generated from exponential family models centered about these values.

A Model for Segmentation and Analysis of Noisy Images

Abstract This article proposes a statistical model for image generation that provides automatic segmentation of images into intensity-differentiated regions and facilitates the quantitative assessment of uncertainty associated with identified image features. The model is specified hierarchically within the Bayesian paradigm. At the lowest level in the hierarchy, a Gibbs distribution is used to specify a probability distribution on the space of all possible partitions of the discretized image scene. An important feature of this distribution is that the number of partitioning elements, or image regions, is not specified a priori. At higher levels in the hierarchical specification, random variables representing emission intensities are associated with regions and pixels. Observations are assumed to be generated from exponential family models centered about these values.

Two-dimensional block Kalman filtering for image restoration

This paper is concerned with developing an efficient two-dimensional (2-D) block Kalman filtering for digital image restoration. A new 2-D multiinput, multioutput (MIMO) state-space structure for modeling the image generation process is introduced. This structure is derived by arranging a vector autoregressive (AR) model with a causal quarter-plane region of support in block form. This model takes into account the correlations of the image data in successive neighboring blocks and, as a result, reduces the edge effects prominent in the available Kalman strip filtering techniques. The degradation model for an infinite extent linear space invariant (LSI) blur and white Gaussian (WG) noise is also modeled by an MIMO block state-space equation stemmed from a single-input single-output (SISO) 2-D state-space structure. The image generation model and the degradation model are combined to yield a composite block-state dynamic structure. The block Kalman filtering equations are obtained for this dynamic structure and then used to compute the suboptimal filter estimates of a noisy and blurred image.

A lens and aperture camera model for synthetic image generation

This paper extends the traditional pin-hole camera projection geometry, used in computer graphics, to a more realistic camera model which approximates the effects of a lens and an aperture function of an actual camera. This model allows the generation of synthetic images which have a depth of field, can be focused on an arbitrary plane, and also permits selective modeling of certain optical characteristics of a lens. The model can be expanded to include motion blur and special effect filters. These capabilities provide additional tools for highlighting important areas of a scene and for portraying certain physical characteristics of an object in an image.

A terrain and cloud computer image generation model

A method is presented that produces a realistic simulation of terrain using source data from the Defense Mapping Agency (DMA). Spatially compact objects are presently modeled with some success at realism. Because terrain is large in extent it exhibits special problems. These problems include: data base size and generation, masking priority, and shading. An array of height values can be created from the DMA terrain file. Height is a single valued function of the array coordinate pair. The pair has implicit positional information. A masking priority algorithm is used which has no restriction on observer geometry. The algorithm searches the height array along the line of sight for the first visible position in the array. This approach requires no clipping or depth sorting of surface representations. The height array indices are used to access the pixel shade from a registered, orthonormal texture array. The use of texture provides a simple way to add nonlinear shading for a realistic representation with flexible data base generation. Atmospheric attenuation is a method used to increase realism. Methods are in use which produce haze or fog with constant attenuation. The line-of-sight search is used to produce realistic clouds with spatially variable attenuation. Terrain and clouds are adequately simulated using this method of modeling. The data structures of height and texture arrays are simple, easily generated, and make good use of storage resources. Resolving masking priority along the line of sight is a visible surface algorithm which has advantages over hidden surface algorithms. Realistic representations such as texture and atmospheric attenuation are used with some success.