CVGIP: Image Understanding最新文献

In this paper we address the problem of methodologies for computer vision. In the first part we will present a brief survey of the Marr paradigm, e.g., what David Marr called his philosophy. We will emphasize the sequence of hypotheses which progressively makes the scene recovery approach explicit as well as the limitations of this approach. We then present the goal-directed approach as an alternative to the recovery school: behaviorism versus reconstructionism. We show that this dichotomy is not the only possible one and introduce the idealism versus empiricism dichotomy. We propose some directions toward a new methodology in a systemic framework involving another, higher-level, methodological dichotomy: systemism versus reductionism. In this new framework we try to exploit of all the sources of constraints, and, thereby, to reconcile some of the previous approaches like recovery school and purposive vision.

In this paper, an algorithm for grouping edges belonging to straight lines is presented. The algorithm uses as input data a labeled set of edge points represented by a list of coordinate-label pairs. The output is a graph whose nodes are rectilinear segments linked by relational properties. Collinearity, convergence, and parallelism can be easily taken into account. The main novelty of the method lies in extending the use of the Hough transform to a symbolic domain (i.e., labeled edges); it is shown that edge labeling can be used to partition the Hough space and to isolate contributions coming from different image areas. Moreover, it is demonstrated that a simple focusing mechanism can be applied (in order to speed up the matching with 3D models) by using relational properties provided by the output graph. In order to confirm the algorithm′s performances, results on synthetic images containing randomly generated textures of straight lines are presented. Finally, a complex road image is considered to point out the advantages of using the proposed representation and the attention-focusing mechanism to solve real-world problems.

Previous work has demonstrated that the task of recovering local disparity measurements can be reduced to measuring the local phase difference between bandpass signals extracted from the left and right images. These earlier techniques were presented as the first simple stages of a more complex stereopsis algorithm. Various local phase difference measurement techniques are examined and a new technique based upon normalized crosscorrelation is presented. A more complete stereopsis algorithm based upon this technique capable of recovering local surface structure from raw image measurements is described. Results obtained with the algorithm are shown on random dot stereopairs, standard stereopairs from the existing literature, and on a calibrated stereopair for which the ground truth is known.

This paper presents an estimation algorithm and error analysis for single linear oriented pattern in images. The estimation is formulated in terms of minimizing an objective function, using the Lagrange multiplier rule. No specific noise model is assumed. The estimation algorithm uses the intensity image of a flow pattern and directly determines a symbolic description of the pattern. No preprocessing or enhancement is needed on the intensity image or any intermediate data. This results in an efficient computational algorithm. We show that it is feasible to directly compute relative divergence, curl, and deformation from the intensity image of an oriented flow pattern. These relative properties are further used for identification of the type of pattern in the intensity image. Since an oriented pattern is corrupted by noise and is distorted to some degree from a linear flow pattern, quality measures of the estimation are proposed. The sampling mean, sampling variance, and energy of noise are computed to characterize its distribution. A closed-form condition number is used to measure the vulnerability of an estimated critical point position to noise perturbation. We show results for several experiments on fluid flow images and wafer defect patterns.

This paper addresses the problem of identifying perceptually significant segments on general planar curvilinear contours. Lacking a formal definition for what constitutes perceptual salience, we develop subjective criteria for evaluating candidate segmentations and formulate corresponding objective measures. An algorithm following these criteria delivers segments with following properties: (1) each segment is well approximated by a circular arc; (2) each pair of segments describe different sections of the contour; and (3) the curve either terminates or changes in orientation and/ or curvature beyond each end of every segment. The result is a description of the contour at multiple scales in terms of circular arcs that may overlap one another.

Computational techniques involving conics are formulated in the framework of projective geometry, and basic notions of projective geometry such as poles, polars, and conjugate pairs are reformulated as "computational procedures" with special emphasis on computational aspects. It is shown that the 3D geometry of three orthogonal lines can be interpreted by computing conics. We then describe an analytical procedure for computing the 3D geometry of a conic of a known shape from its projection. Real image examples are also given.

This paper presents a new methodology of representing and matching nonplanar developable surfaces (NPDS) for the purpose of three-dimensional objects recognition. The constant-ratio property, a special property of an NPDS, is presented. Using this property, congruence conditions for two NPDS segments are derived that depend only on geometric and numerically computable properties of a surface. Based on these theoretical results, a developable surface description is presented and an algorithm is developed that matches an unknown NPDS with surface models to identify the unknown surface. The practical feasibility of this methodology is studied and is illustrated by various concrete examples using range images. The numerical computation involved in and the noise sensitivity of the approach are also addressed.

Conventional structural pattern recognition methods that rely on thinning approximate the skeleton by polygons and proceed from that step to recognition. Polygonal approximations have the disadvantage that they introduce significant ambiguities. In particular, vertices may correspond either to real corners or to approximations of smooth arcs. We propose a method that relies on topographic features and, in particular, ridge lines. The information about ridge line directions obtained from the underlying surface of the gray tone is used to discriminate between arcs and straight lines. Normally, ridge lines are centered within character strokes, forming skeleton-like ribbons with generally no more than three pixels in width. For each ridge pixel, the tangent direction of the ridge line at the pixel is calculated. These computed tangent directions are then used in the detection of sharp corners and junctions, in the line tracking process, and in the feature decomposition process. Decomposition is achieved using curvature primitives and singular points. The result of the method is a relational feature graph which gives a compact and flexible description of the shapes of the objects in the input image. By not using a conventional thinning algorithm and performing arc and straight line decomposition without a usual polygonal approximation step, our method is able to reduce some artifacts of conventional thinning and to eliminate completely the ambiguities resulting from polygonal approximations.

This paper sets out to study the effect of digitization on curvature. It is put forward that digital curvature is an estimate with accuracy and precision. A theoretical analysis of the curvature estimation problem identifies the possible sources of errors in digital curvature estimation. In the literature, five essentially different classes of methods are found. From theoretical analysis of the methods, as well as by random experiments on generated arcs, we establish that almost all existing methods suffer from a severe directional inaccuracy and/or poor precision. Errors depend on the method, orientation, and scale, ranging from 1% to over 1000%. We present a practical solution with a residual error between 1% and 60%, also giving recommendations for the required resolution of the image.