International Journal on Artificial Intelligence Tools

The concentration of the cones and ganglion cells is much higher in the fovea than the rest of the retina. This non-uniform sampling results in a retinal image that is sharp at the fixation point, where a person is looking, and blurred away from it. This difference between the sampling rates at the different spatial locations presents us with the question of whether we can employ this biological characteristic to achieve better image compression. This can be achieved by compressing an image less at the fixation point and more away from it. It is, however, known that the vision system employs more that one fixation to look at a single scene which presents us with the problem of combining images pertaining to the same scene but exhibiting different spatial contrasts. This article presents an algorithm to combine such a series of images by using image fusion in the gradient domain. The advantage of the algorithm is that unlike other algorithms that compress the image in the spatial domain our algorithm results in no artifacts. The algorithm is based on two steps, in the first we modify the gradients of an image based on a limited number of fixations and in the second we integrate the modified gradient. Results based on measured and predicted fixations verify our approach.

In this paper, we describe a fast and efficient method for multi-modal and discontinuity-preserving image registration, implemented on graphics hardware. Multi-sensory data fusion and medical image analysis often pose the challenging task of aligning dense, non-rigid and multi-modal images. However, also optical sequences or stereo image pairs may present variable illumination conditions and noise. The above problems can be addressed by an invariant similarity measure, such as mutual information. Additionally, when using a regularized approach to deal with the ill-posedness of the problem, one has to take care of preserving discontinuities at the motion boundaries. Our approach efficiently addresses the above issues through a primal-dual convex estimation framework, using an approximated Hessian matrix that decouples pixel dependencies, while being asymptotically correct. At the same time, we achieve a high computational efficiency by means of pre-quantized kernel density estimation and differentiation, as well as a parallel implementation on the GPU. Our approach is demonstrated on ground-truth data from the Middlebury database, as well as medical and visible-infrared image pairs.

In this paper, a steerable, interactive projection display that has the shape of a disk is presented. Interactivity is provided through sensitivity to the contact of multiple fingertips and is achieved through the use of a RGBD camera. The surface is mounted on two gimbals which, in turn, provide two rotational degrees of freedom. Modulation of surface posture supports the ergonomy of the device but can be, alternatively, used as a means of user-interface input. The geometry for mapping visual content and localizing fingertip contacts upon this steerable display is provided, along with pertinent calibration methods for the proposed system. An accurate technique for touch detection is proposed, while touch detection and projection accuracy issues are studied and evaluated through extensive experimentation. Most importantly, the system is thoroughly evaluated as to its usability, through a pilot application that was developed for this purpose. We show that the outcome meets real-time performance, accuracy and usability requirements for employing the approach in human computer interaction.

We show that the results from searching 2D images or a video sequence, with a 3D head model can be improved by using detailed sub-models. These parts are initialised with the full model result and are allowed to search independently of that model, and of each other, using the same algorithm. The final results for the sub-models can be reported exactly, or optionally fed back into the full model to be constrained by its parameter space. In the case of a video sequence this can then be used in the initialisation of the next frame. We tested various data sets, constrained and unconstrained, including a variety of lighting conditions, poses, and expressions. Our investigation showed that using the sub-models improved on the original full model result on all but one of the data sets.

Gender recognition from facial images plays an important role in biometric applications. Employing Dyadic wavelet Transform (DyWT) and Local Binary Pattern (LBP), we propose a new feature descriptor DyWT-LBP for gender recognition. DyWT is a multi-scale image transformation technique that decomposes an image into a number of sub-bands which separate the features at different scales. DyWT is a kind of translation invariant wavelet transform that has a better potential for detection than Discrete Wavelet Transform (DWT). On the other hand, LBP is a texture descriptor and is known to be the best for representing texture micro-patterns, which play a key role in the discrimination of different objects in an image. For DyWT, we used spline dyadic wavelets (SDW). There exist many types of SDW; we investigated a number of SDWs for finding the best SDW for gender recognition. The dimension of the feature space generated by DyWT-LBP descriptor becomes excessively high. To tackle this problem, we apply a feature subset selection (FSS) technique that not only reduces the number of features significantly but also improves the recognition accuracy. Through a large number of experiments performed on FERET and Multi-PIE databases, we report for DyWT-LBP descriptor the parameter settings, which result in the best accuracy. The proposed system outperforms the stat of the art gender recognition approaches; it achieved a recognition rate of 99.25% and 99.09% on FERET and Multi-PIE databases, respectively.

Spectral rendering, or the synthesis of images by taking into account the constituent wavelengths of white light, enables the rendering of iridescent colors caused by phenomena such as dispersion, diffraction, interference and scattering. Caustics, the focusing and defocusing of light through a refractive medium, can be interpreted as a special case of dispersion where all the wavelengths travel along the same paths. In this paper we extend Adaptive Caustic Mapping (ACM), a previously proposed caustics mapping algorithm, to handle physically-based dispersion. Because ACM can display caustics in real-time, it is amenable to extension to handle the more general case of dispersion. We also present a novel algorithm for filling in the gaps that occur due to discrete sampling of the spectrum. Our proposed method runs in screen-space, and is fast enough to display plausible dispersion phenomena at real-time and interactive frame rates.

We present a novel way to efficiently compute Riemannian geodesic distance over a two- or three-dimensional domain. It is based on a previously presented method for computation of geodesic distances on surface meshes. Our method is adapted for rectangular grids, equipped with a variable anisotropic metric tensor. Processing and visualization of such tensor fields is common in certain applications, for instance structure tensor fields in image analysis and diffusion tensor fields in medical imaging.

The included benchmark study shows that our method provides significantly better results in anisotropic regions in 2-D and 3-D and is faster than current stat-of-the-art solvers in 2-D grids. Additionally, our method is straightforward to code; the test implementation is less than 150 lines of C++ code. The paper is an extension of a previously presented conference paper and includes new sections on 3-D grids in particular.

The development of vision-based human activity recognition and analysis systems has been a matter of great interest to both the research community and practitioners during the last 20 years. Traditional methods that require a human operator watching raw video streams are nowadays deemed as at least ineffective and expensive. New, smart solutions in automatic surveillance and monitoring have emerged, propelled by significant technological advances in the fields of image processing, artificial intelligence, electronics and optics, embedded computing and networking, molding the future of several applications that can benefit from them, like security and healthcare. The main motivation behind it is to exploit the highly informative visual data captured by cameras and perform high-level inference in an automatic, ubiquitous and unobtrusive manner, so as to aid human operators, or even replace them. This survey attempts to comprehensively review the current research and development on vision-based human activity recognition. Synopses from various methodologies are presented in an effort to garner the advantages and shortcomings of the most recent state-of-the-art technologies. Also a first-level self-evaluation of methodologies is also proposed, which incorporates a set of significant features that best describe the most important aspects of each methodology in terms of operation, performance and others and weighted by their importance. The purpose of this study is to serve as a reference for further research and evaluation to raise thoughts and discussions for future improvements of each methodology towards maturity and usefulness.