One of the most basic characteristics of the image is accompanied by its blur. It was 1962 that I had discovered for the first time in the world that the blur was a Gaussian type. In this paper the outline is described about historical details concerning this circumstances.

The joint histogram of second order scale space differential invariants of natural images (including textures) is typically clustered about parabolic surface patches, whereas symmetrical elliptical patches (local convexities or concavities) are very rare and symmetrical hyperbolical patches also occur less frequently than parabolic patches. We trace the origin of this striking effect in the context of Gaussian random noise. For this case one may derive the joint histogram of curvedness and shape index analytically. The empirical observations are fully corroborated. In deriving these results we introduce a polar coordinate system in the space of second order scale space derivatives that turns out to be particularly useful in the study of the statistics of local curvature properties. The empirical observations apply also to non-Gaussian noise (e.g., Brownian noise) as well as to photographs of natural scenes. We discuss general arguments that help explain these observations.

We consider the classification problem as a problem of approximation of a given training set. This approximation is constructed in a multiresolution framework, and organized in a tree-structure. It allows efficient training and query, both in constant time per training point. The proposed method is efficient for low-dimensional classification and regression estimation problems with large data sets.

Recent advances of Web information systems such as e-commerce and e-learning have created very large but hidden demands on conceptual multiresolution analysis for more generalized information analysis, cognition and modeling. To meet the demands in a general way, its modeling is formulated based on modern algebraic topology. To be specific, the modeling formulation is worked out in an incrementally modular abstraction hierarchy with emphasis on the two levels of the hierarchy appropriate for conceptual modeling: the adjunction space level and the cellular structured space level. Examples are shown to demonstrate the usefulness of the presented model as well as an implementation of a flower structure case.

Computation of scale space images requires numerical integration of partial differential equations, which demands large computational costs especially in nonlinear cases. In this paper, we present a computational scheme for nonlinear scale spaces based on iterated filtering of original images. This scheme is found to be a special case of numerical integration with a particularly adapted integration steplength. We show the stability of the iteration with local windows and that with global ones and analyze the deformation of edge waveforms in the filtering. Computational costs are evaluated experimentally for both local and global windows and finally we apply the nonlinear multi-scale smoothing to contrast enhancement of images.

We introduce a crystalline flow for a contour figure analysis. The crystalline flow is a special family of evolving polygons, and is considered as a discrete version of a classical curvature flow. In the evolving process of the crystalline flow, each facet moves toward its normal direction. The velocity of the facet is determined by the nonlocal curvature, which depends on the length of the facet. Different from a classical curvature flow, it is easy to track each facet in a given contour through the evolving process, because a given polygon remains polygonal. This aspect helps us to make a scale-space representation of a contour in an image. In this article, we present a method for extracting dominant corners using a crystalline flow. Experimental results show that our method extracts several sets of dominant corner facets successfully from a given contour figure.

In medical diagnosis, cone beam CT increases the dose absorbed by a patient. However, the radiographic noise (such as quantum noise) in a CT image increases when radiation exposure is reduced. In this paper, we propose a method to improve the CT image degraded by the quantum mottle based on 2-D wavelet transform modulus sum (WTMS). The noise and regular parts of an image can be observed by tracing the evolution of its 2-D WTMS across scales. Our experimental results show that most of the quantum mottle in the 2-D projections is removed by the proposed method and the edges preserved well. We investigate the relation between the number of X-ray photons and the quality of the denoised images. The result shows the possibility that a patient's dose can be reduced about 10% with the same visual quality by our method.

We describe a multiresolution 3D active balloon model to trace the boundaries of moving object. This model is able to analyze a shape hierarchically using 3D scale-space. The 3D scale-space can be determined by changing the parameters of the active balloon. We extended 2D process-grammar to describe the deformation process between a shape and a sphere, based on topological scale-space analysis. The geometric invariant features were used to analyze the deformation of nonrigid shapes. We analyzed the motion of a heart by using MRI data.

In the image recognition problem, it is very important how we represent the image. Considering this, we propose a new representational method of images based on the stability in scale-space. In our method, the image is segmented and represented as a hierarchical region graph in scale-space. The object is represented as feature graph, which is subgraph of region graph. In detail, the region graph is defined on the image with the relation of each segment hierarchically. And the feature graph is determined based on the "life-time" of the graph of the object in scale-space. This "life-time" means how long feature graph lives when the scale parameter is increased. We apply our method to the face detection problem, which is foundmental and difficult problem in face recognition. We determine the feature graph of the frontal human face statistical point of view. We also build the face detection system using this feature graph to show how our method works efficiently.

Probabilistic inference by means of a massive probabilistic model usually has exponential-order computational complexity. For such massive probabilistic model, loopy belief propagation was proposed as a scheme to obtain the approximate inference. It is known that the generalized loopy belief propagation is constructed by using a cluster variation method. However, it is difficult to calculate the correlation in every pair of nodes which are not connected directly to each other by means of the generalized loopy belief propagation. In the present paper, we propose a general scheme for calculating an approximate correlation in every pair of nodes in a probabilistic model for probabilistic inference. The general scheme is formulated by combining a cluster variation method with a linear response theory.

This paper presents a new asynchronous multiplier. The original array structure is divided into two asymmetric arrays, called an upper array and a lower array. For the lower array, Left to Right scheme is applied to take advantage of a fast computation and low power consumption as well. Simulation results show that the proposed multiplier has 40% of performance improvement with a relatively lower power consumption. The multiplier has been implemented in a CMOS 0.35 µm technology and proved functionally correct.

Many application programs can be modeled as a subgraph isomorphism problem. However, this problem is generally NP-complete and difficult to compute. A custom computing circuit is a prospective solution for such problems. This paper examines various accelerator designs for subgraph isomorphism problems based on Ullmann's algorithm and Konishi's algorithm. These designs are quantitatively evaluated from two points of view: logic scale and execution time. Our study revealed that Ullmann's design is faster but larger in logic scale. Partially sequential versions of Ullmann's algorithm can be more cost-effective than Ullmann's original design. The hardware of Konishi's algorithm is smaller in logic scale, operates at a higher frequency, and is more cost-effective.

Recently, layer-4 (L4) switches have been widely used as load balancing front-end routers for Web server clusters. The typical L4 switch attempts to balance load among the servers by estimating load using the load metrics measured in the front-end and/or the servers. However, insufficient load metrics, measurement overhead, and feedback delay often cause misestimate of server load. This may incur significant dynamic load imbalance among the servers particularly when the variation of requested content is high. In this paper, we propose a new content sniffer based load distribution strategy. By sniffing the requests being forwarded to the servers and by extracting load metrics from them, the L4 switch with our strategy more timely and accurately estimates server load without the help of back-end servers. Thus it can properly react to dynamic load imbalance among the servers under various workloads. Our experimental results demonstrate substantial performance improvements over other load balancing strategies used in the typical L4 switch.

Many fuzzy traffic controllers adjust the extension time of the green phase with the fuzzy input variables, arrival and queue. However, in our experiments, we found that the two input variables are not sufficient for an intersection where traffic flow rates change and thus, in this paper, traffic volume is used as an additional variable. Traffic volume is defined as the number of vehicles entering an intersection every second. In designing a fuzzy traffic controller, an ad-hoc approach is usually used to find membership functions and fuzzy control rules showing good performance. That is, initial ones are generated by human operators and modified many times based on the results of simulation. To partially overcome the limitations of the ad-hoc approach, we use genetic algorithms to automatically determine the membership functions for terms of each fuzzy variable when fuzzy control rules are given by hand. The experimental results indicate that a fuzzy logic controller with volume variable outperforms conventional ones with no volume variable in terms of the average delay and the average velocity. Also, the controller shows better performance when membership functions generated by a genetic algorithms instead of ones generated by hand are used.

A set of systematic experiments on intelligent email categorization has been conducted with different machine learning algorithms applied to different parts of data in order to achieve the most correct classification. The categorization is based on not only the body but also the header of an email message. The metadata (e.g. sender name, sender organization, etc.) provide additional information that can be exploited to improve the categorization capability. Results of experiments on real email data demonstrate the feasibility of our approach to find the best learning algorithm and the metadata to be used, which is a very significant contribution in email classification. It is also shown that categorization based only on the header information is comparable or superior to that based on all the information in a message for all the learning algorithms considered.

A wide variety of visual textures could be successfully modeled as spatially variant by quantitatively describing them through the variation of their local spatial frequency and/or local orientation components. This class of patterns includes flow-like, granular or oriented textures. Modeling is achieved by assuming that locally, textured images contain a single dominant component describing their local spatial frequency and modulating amplitude or contrast. Spatially variant textures are non-homogeneous in the sense of having nonstationary local spectra, while remaining locally coherent. Segmenting spatially variant textures is the challenging task undertaken in this paper. Usually, the goal of texture segmentation is to split an image into regions with homogeneous textural properties. However, in the case of image regions with spatially variant textures, there is no global homogeneity present and thus segmentation passes through identification of regions with globally nonstationary, but locally coherent, textural content. Local spatial frequency components are accurately estimated using Gabor wavelet outputs along with the absolute magnitude of the convolution of the input image with the first derivatives of the underlying Gabor function. In this paper, a frequency estimation approach is used for segmentation. Indeed, at the boundary between adjacent textures, discontinuities occur in texture local spatial frequency components. These discontinuities are interpreted as corresponding to texture boundaries. Experimental results are in remarkable agreement with human visual perception, and demonstrate the effectiveness of the proposed technique.

Recently, several promising multiobjective evolutionary algorithms such as PESA, NSGA-II, and SPEA2 have been developed. In this paper, we also propose a new multiobjective evolutionary algorithm whose performance is comparable to or better than those promising algorithms. In the new algorithm proposed here, an age concept is introduced and utilized to make the efficiency of the offspring generation high. The performance of the proposed algorithm is superior to those of the promising algorithms mentioned above for several test functions. In this paper, the proposed algorithm will be explained only in two dimensional parameter and objective space to show manifestly the meaning of an age concept.

This letter presents the implementation results of an application-level cache file system, _MCFS, which is specifically designed to provide efficient caching and transmission mechanisms for streaming media. The file system is built on a virtual file disk which is constructed as a single large file on a general-purpose file system. _MCFS suits the access requirement of continuous media caching and provides an efficient I/O mechanism for cache servers. The experimental results show that _MCFS outperforms the comparison model and provides a consistent I/O bandwidth.

Adapting to the structure of 2-D H-Transform, this paper proposes a novel wavelet domain half-pixel motion compensation algorithm HMRME (Half-pixel Multi-Resolution Motion Estimation). The primary objective of this study is the reduction of the aliasing effect caused by the down-sampling in the wavelet transform under the complexity constraints. The conventional multi-resolution motion estimation scheme can be combined with the half-pixel interpolation method to generate a new high-performance wavelet video codec. The preliminary results show that the performance of HMRME rises above its counterparts, the Multi-Resolution Motion Estimation (MRME) and the Adaptive Multi-Resolution Motion Estimation (AMRME).

This Letter presents a fast codeword search algorithm based on ordered Hadamard transform. Before encoding, the ordered Hadamard transform is performed offline on all codewords. During the encoding process, the ordered Hadamard transform is first performed on the input vector, and then a new inequality based on characteristic values of transformed vectors is used to reject the unlikely transformed codewords. Experimental results show that the algorithm outperforms many newly presented algorithms in the case of high dimensionality, especially for high-detail images.

This paper presents an edge alignment method for stitching images when they have large displacements and light changes. First, without building any correspondences, the proposed method predicts all possible translation solutions by examining the consistency between edge positions. Then, the best solution can be obtained from the set of possible translations by a verification process. The proposed method has better capabilities to stitch images when they have large light changes and displacements. Since the method doesn't require building any correspondences or involve any optimization process, it performs more efficiently than other correlation techniques like feature-matching or phase-correlation approaches. Due to its simplicity and efficiency, different images can be very quickly aligned (less than 0.1 seconds) with the proposed method. Experimental results are provided to verify the superiority of the proposed method.