Graphical Models and Image Processing最新文献

To achieve very low bit-rate transmission, many video coding papers suggest extraction of stationary objects and the background and coding only of moving objects. These techniques, which could seriously distort the output, are mostly applied to spatial domain. A novel scalable adaptive temporal segmentation algorithm for video coding is proposed in this paper. Our proposed algorithm is scalable in terms of compression, quality, and some specific features. A multifunction control of the output is formulated as a cost function, which is flexible enough to implement any desirable criteria, such as channel capacity or reconstruction quality. Our algorithm uses a simple but efficient quantization process instead of using a computationally expensive discontinuity detection. The approach can also be regarded as a combination of adaptive temporal decimation and segmented video coding. The overall bit rate can be regulated according to the user's specifications. Experimental results show that the overall bit rate of video coding can be controlled easily using our scheme. The output quality is good for both subjective and objective tests. We have also found that the size of the time window (N) has significant effects on the quality of the reconstruction, while our proposed algorithm outperforms the adaptive temporal decimation algorithm.

The problem of estimating the center and radius of a circle from a given set of noisy coordinate measurements has many applications. Even though the estimation process is inherently nonlinear, it is possible to obtain linear estimators, which are attractive because of their simplicity. This paper presents an approximate maximum likelihood estimator of circle parameters. It is linear. Simulation results showed that under conditions of small arc length and high noise, which occur often in practice, the new estimator outperforms the linear estimator of S. Thomas and Y. T. Chan (Comput. Vision Graph. Image Process.45 (1989), 362–370).

By asegmentedimage, we mean a digital image in which each point is assigned a unique label that indicates the object to which it belongs. By the foreground (objects) of a segmented image, we mean the objects whose properties we want to analyze and by the background, all the other objects of a digital image. If one adjacency relation is used for the foreground of a 3D segmented image (e.g., 6-adjacency) and a different relation for the background (e.g., 26-adjacency), then interchanging the foreground and the background can change the connected components of the digital picture. Hence, the choice of foreground and background is critical for the results of the subsequent analysis (like object grouping), especially in cases where it is not clear at the beginning of the analysis what constitutes the foreground and what the background, since this choice immediately determines the connected components of the digital picture. A special class of segmented digital 3D pictures called “well-composed pictures” will be defined. Well-composed pictures have very nice topical and geometrical properties; in particular, the boundary of every connected component is a Jordan surface and there is only one type of connected component in a well-composed picture, since 6-, 14-, 18-, and 26-connected components are equal. This implies that for a well-composed digital picture, the choice of the foreground and the background is not critical for the results of the subsequent analysis. Moreover, a very natural definition of a continuous analog for well-composed digital pictures leads to regular properties of surfaces. This allows us to give a simple proof of a digital version of the 3D Jordan–Brouwer separation theorem.

We show that a rotation in three dimensions can be achieved by a composition of three shears, the first and third along a specified axis and the second along another given axis orthogonal to the first; this process is invertible. The resulting rotation algorithm is practical for the processing of fine-grained digital images, and is well adapted to the access constraints of common storage media such as dynamicRAMor magnetic disk. For a 2-D image, rotation by composition of three shears is well known. For 3-D, an obvious nine-shear decomposition has been mentioned in the literature. Our three-shear decomposition is a sizable improvement over that, and is the best that can be attained—just two shears won't do. Also, we give a brief summary of how the present three-shear decomposition approach generalizes to any linear transformations of unit determinant in any number of dimensions.

When processing shape information derived from a noisy source such as a digital scanner, it is often useful to construct polygonal or curved outlines that match the input to within a specified tolerance and maximize some intuitive notions of smoothness, simplicity, and best fit. The outline description should also be concise enough to be competitive with binary image compression schemes. Otherwise, there will be a strong temptation to lose the advantages of the outline representation by converting back to a binary image format. This paper proposes a two-stage pipeline that provides separate control over the twin goals of smoothness and conciseness: the first stage produces a specification for a set of closed curves that minimize the number of inflections subject to a specified error bound; the second stage produces polygonal outlines that obey the specifications, have vertices on a given grid, and have nearly the minimum possible number of vertices. Both algorithms are reasonably fast in practice, and can be implemented largely with low-precision integer arithmetic.

In this note, an efficient class of alternating sequential filters (ASFs) in mathematical morphology is presented to reduce the computational complexity in the conventional ASFs about a half. The performance boundary curves of the new filters are provided. Experimental results from applying these new ASFs to texture classification and image filtering (grayscale and binary) show that comparable performance can be achieved while much of the computational complexity is reduced.

The number of nodes of an edge quadtree is the measure of its space complexity. This number depends on the figure's shape, its resolution and its precision. The goal of this work is to prove that a relation exists between the number of nodes of an edge-quadtree and these three parameters. To reach this goal an experimental approach has been used. A unique value to represent both the resolution and the precision is used. To measure the shape of the image we use the fractal dimension. A methodology to calculate the fractal dimension and the fractal measure is proposed. These three parameters being given, we use a neural network to approximate the sought function. The computational results show the effectiveness of this approach.

The three-dimensional Cartesian geometric moments (for short 3-D moments) are important features for 3-D object recognition and shape description. To calculate the moments of objects in a 3-D image by a straightforward method requires a large number of computing operations. Some authors have proposed fast algorithms to compute the 3-D moments. However, the problem of computation has not been well solved since all known methods require computations of orderN³, assuming that the object is represented by anN×N×Nvoxel image. In this paper, we present a discrete divergence theorem which can be used to compute the sum of a function over ann-dimensional discrete region by a summation over the discrete surface enclosing the region. As its corollaries, we give a discrete Gauss's theorem for 3-D discrete objects and a discrete Green's theorem for 2-D discrete objects. Using a fast surface tracking algorithm and the discrete Gauss's theorem, we design a new method to compute the geometric moments of 3-D binary objects as observed in 3-D discrete images. This method reduces the computational complexity significantly, requiring computation ofO(N²). This reduction is demonstrated experimentally on two 3-D objects. We also generalize the method to deal with high-dimensional images. Some 3-D moment invariants and shape features are also discussed.

Ordered dither (C. N. Judice, J. F. Jarvis, and W. H. Ninke, inProceedings of the S.I.D. 15, 4(Fourth Quarter),1974,pp. 161–169.) is an efficient and widely used halftoning technique. One disadvantage with this method is that it produces blurred images. Recently, a space-filling curve ordered dithering approach (Y. Zhang, submitted for publication.) has been proposed that improves the ordered dither in image quality by generating clustered dot patterns along a space-filling curve over the image. In this method the cluster size is fixed in the process of halftoning. This paper presents a new ordered dithering method, called adaptive ordered dither, that does halftoning by using a space-filling curve to perform an adaptive variation of the cluster size. Experimental results demonstrate that the new method can significantly improve the space-filling curve ordered dither in revealing image details.