Pattern Recognition Letters Research Articles

System-on-programmable-chip implementation for on-line face recognition

In this paper, the design of a parallel architecture for on-line face recognition using weighted modular principal component analysis (WMPCA) and its system-on-programmable-chip (SoPC) implementation are discussed. The WMPCA methodology, proposed by us earlier, is based on the assumption that the rates of variation of the different regions of a face are different due to variations in expression and illumination. Given a database of sample faces for training and a query face for recognizing, the WMPCA methodology involves division of the face into horizontal regions. Each of these regions are analyzed independently by computing the eigenfeatures and comparing the same with the corresponding eigenfeatures of the faces stored in the sample database to calculate the corresponding error. The final decision of the face recognizer is based on the weighted sum of the errors computed from each of the regions. These weights are calculated based on the extent to which the various samples of the subject are spread in the eigenspace. The WMPCA methodology has a better recognition rate compared to the modular PCA approach developed by Rajkiran and Vijayan [Rajkiran, G., Vijayan, K., 2004. An improved face recognition technique based on modular PCA approach. Pattern Recognition Letters, 25(4), 429–436]. The methodology also has a wide scope for parallelism. We present an architecture that exploits this parallelism and implement the same as a system-on-programmable-chip on an ALTERA based field programmable gate array (FPGA) platform. The implementation has achieved a processing speed of about 26 frames per second at an operating frequency of 33.33 MHz.

Detecting and removing specularities in facial images

Specularities often confound algorithms designed to solve computer vision tasks such as image segmentation, object detection, and tracking. These tasks usually require color image segmentation to partition an image into regions, where each region corresponds to a particular material. Due to discontinuities resulting from shadows and specularities, a single material is often segmented into several sub-regions. In this paper, a specularity detection and removal technique is proposed that requires no camera calibration or other a priori information regarding the scene. The approach specifically addresses detecting and removing specularities in facial images. The image is first processed by the Luminance Multi-Scale Retinex [B.V. Funt, K. Barnard, M. Brockington, V. Cardei, Luminance-Based Multi-Scale Retinex, AIC’97, Kyoto, Japan, May 1997]. Second, potential specularities are detected and a wavefront is generated outwards from the peak of the specularity to its boundary or until a material boundary has been reached. Upon attaining the specularity boundary, the wavefront contracts inwards while coloring in the specularity until the latter no longer exists. The third step is discussed in a companion paper [M.D. Levine, J. Bhattacharyya, Removing shadows, Pattern Recognition Letters, 26 (2005) 251–265] where a method for detecting and removing shadows has also been introduced. The approach involves training Support Vector Machines to identify shadow boundaries based on their boundary properties. The latter are used to identify shadowed regions in the image and then assign to them the color of non-shadow neighbors of the same material as the shadow. Based on these three steps, we show that more meaningful color image segmentations can be achieved by compensating for illumination using the Illumination Compensation Method proposed in this paper. It is also demonstrated that the accuracy of facial skin detection improves significantly when this illumination compensation approach is used. Finally, we show how illumination compensation can increase the accuracy of face recognition.

Note on the equivalence relationship between Renyi-entropy based and Tsallis-entropy based image thresholding

Renyi entropy based image thresholding in [Sahoo, P., Wilkins, C., et al., 1997. Thresholding selection using Renyi’s entropy. Pattern Recognition 30 (1), 71–84; Sahoo, P.K., Arora, G., 2004. A thresholding method based on two dimensional Renyi’s entropy. Pattern Recognition 37 (6), 1149–1161] and Tsallis entropy based image thresholding in [Albuquerque, M.P., Esquef, I.A., et al., 2004. Image thresholding using Tsallis entropy. Pattern Recognition Letters 25 (9), 1059–1065] are two important global threshold selection approaches in image segmentation. In this correspondence, the equivalence relationship between these two approaches is revealed, that is to say, with the same parameter, these two approaches will obtain the same threshold.

RETRACTED: Invariance image analysis using modified Zernike moments

This article has been retracted at the request of the Editors-in-Chief. Please see http://www.elsevier.com/locate/withdrawalpolicy . Reason: This article is retracted at the request of the Editors-in-Chief and agreed with the authors. It has come to the attention of the Editors-in-Chief of Pattern Recognition Letters (PRL) that the article “Invariance image analysis using modified Zernike moments”, N.K. Kamila, S. Mahapatra, S. Nanda, Pattern Recognition Letters 26 (2005)747–753 closely resembles “Invariance analysis of improved Zernike moments”, Y. Bin and P. Jia-Xiong, J. Optics A: Pure Appli. Optics 4 (2002) 606–614. In the Kamila et al. article the main part of the text and the experiment presentation (though with different data) are almost a literal reproduction of those appearing in the latter article, and where the text is verbatim, the passage is not in inverted commas and the original source is not quoted. PRL author guidelines clearly state that consideration of an article for publication in the Journal is made on the basis that the article is original, unpublished, and not being considered for publication elsewhere.

A note on core research issues for statistical pattern recognition

Is statistical pattern recognition a collection of topics and techniques borrowed from other fields? Or has it an identity of its own, generating specific research questions, producing its own theoretical framework? This question arises occasionally between participants of the workshops and conferences focusing on the statistically oriented procedures and applications in pattern recognition. Although many pattern recognition scientists seem pretty sure about the cultural identity of their research field, the authors think that the recent developments of closely related disciplines (e.g., machine learning and neural networks) and the increasing number of research issues that pattern recognition shares with such disciplines makes the answer to the above question more difficult than in past times. Therefore, this paper aims to stimulate discussion in the pattern recognition community on the structural differences between statistical pattern recognition and closely related disciplines in order to clarify the present cultural identity and the core research issues. According to computer science encyclopaedia and some famous textbooks, pattern recognition (PR) can be defined as the discipline that studies theories and methods for designing machines that are able to recognise patterns in noisy data (Srihari and Govindaraju, 1993; Duda et al., 2001; Fukunaga, 1990; Devijver and Kittler, 1982). Although this is only one of the possible definitions for PR, it points out well the ‘‘engineering’’ nature of this scientific discipline, because it says that the final goal of PR is the design of ‘‘machines’’. In this engineering perspective on PR, it is quite obvious that PR theory has multi-disciplinary roots, because engineering disciplines aim to bridge the gap between real-world applications and the so-called pure disciplines (mathematics, statistics, physics, etc.). It is worth noting that this multi-disciplinary character of the theoretical foundations of PR was already evident in the seminal book by Watanabe (1985). Watanabe pointed out that pattern recognition can be regarded as a statistical decisionmaking task, as a structure analysis problem, as an induction process, as a perception task, etc., thereby acknowledging the multiple cultural roots of PR. With regard to the strong relations between PR and artificial intelligence, Tveter recently stated that many artificial intelligence methods can be Pattern Recognition Letters 23 (2002) 493–499

Combination of neural and statistical algorithms for supervised classification of remote-sensing images

Various experimental comparisons of algorithms for supervised classification of remote-sensing images have been reported in the literature. Among others, a comparison of neural and statistical classifiers has previously been made by the authors in (Serpico, S.B., Bruzzone, L., Roli, F., 1996. Pattern Recognition Letters 17, 1331–1341). Results of reported experiments have clearly shown that the superiority of one algorithm over another cannot be claimed. In addition, they have pointed out that statistical and neural algorithms often require expensive design phases to attain high classification accuracy. In this paper, the combination of neural and statistical algorithms is proposed as a method to obtain high accuracy values after much shorter design phases and to improve the accuracy–rejection tradeoff over those allowed by single algorithms.

On the intimate relationship between the principle of duality and the Hough transform

This paper shows the relationship between the principle of duality and the Hough transform. We show that the definition of the Hough transform actually corresponds to an application of the principle of duality. In this interpretation, the general mathematical concepts of figures and high–dimensional coordinates in projective space are re–expressed in terms of shapes and mappings used for shape extraction in Euclidean space. So far, some works have demonstrated the equivalence between the definition of several concepts used in pattern matching and the Hough transform. These works have related the Hough transform to particular developments of diverse concepts including template matching, the Radon transform, maximum likelihood estimation and robust statistics. The correspondence between the Hough transform definition and the principle of duality here developed suggests a more general definition of the Hough transform with a deeper meaning in pure geometry. This definition introduces the formalism of projective geometry to shape extraction and analysis and thus the ideas, properties and geometric relationships in the projective space can have an interpretation for the development of pattern matching techniques. We discuss the analogy between the definition of a figure in geometry and the notion of a shape in the Hough transform, and how this relationship and the generalization of the principle of duality to high–dimensional spaces, defined by space coordinates, are related to the extensions of the Hough transform. Parametric forms are used to develop a dual analytic expression of general forms. The intimate relationship between the Hough transform and the principle of duality increases our understanding of the dual nature of pattern matching which can now benefit from established results in projective geometry.

A modified Hausdorff distance between fuzzy sets

In an earlier paper (B.B. Chaudhuri, A. Rosenfeld, Pattern Recognition Letters 17 (1996) 1157–1160) we defined a Hausdorff-like metric distance between fuzzy sets. This distance contained two terms, one representing geometric distance and the other representing dissimilarity. Combining two terms representing two unrelated notions is not very appealing. In this paper, a modified metric is proposed which contains a single term representing geometric distance only; which reduces to the classical Hausdorff distance when the fuzzy sets are crisp; and which is relatively insensitive to noise.

A neural network approach to quality control of padlock manufacturing : Hakulinen, A.; Hakkarainen, J. Pattern Recognition Letters, Vol. 17, No. 4, pp. 357–362 (1996)

A neural network approach to quality control of padlock manufacturing : Hakulinen, A.; Hakkarainen, J. Pattern Recognition Letters, Vol. 17, No. 4, pp. 357–362 (1996)

Material characterization by ultrasonics using unsupervised competitive learning : Chou, C.P.; Ho, B.; Sheu, J.T. Pattern Recognition Letters, Vol. 16, No. 7, pp. 769–777 (Jul. 1995)

Material characterization by ultrasonics using unsupervised competitive learning : Chou, C.P.; Ho, B.; Sheu, J.T. Pattern Recognition Letters, Vol. 16, No. 7, pp. 769–777 (Jul. 1995)

Edge linking by using causal neighborhood window

Xie, M., Edge linking by using causal neighborhood window, Pattern Recognition Letters 13 (1992) 647-656. A novel edge linking algorithm will be presented in this paper. We shall make use of two concepts: (a) the concept of horizontal edge element and (b) the concept of causal ne,~hborhood window for realizing the edge linking. On the basis of these two concepts, the proposed algorithm will carry out (in one pass on the edge map) the edge linking by performing two operations which are the contour chain creation and the contour cha#! extension. For the illustration of the applicability of our algorithm, some experimental results with real images will be shown. Kt:rwords. Edge linking.

A comparison of splitting methods for the identification of corner-points

Espelid, R. and I. Jonassen, A comparison of splitting methods for the identification of corner-points, Pattern Recognition Letters 12 (1991) 79–83. The problem of detecting corner-points in two-dimensional image curves, using splitting methods is addressed. The question of interest is how to reduce computing time without sacrificing quality in corner-point positioning. Three algorithms are compared: the classical Iterative-End-Point algorithm; a modified, line search method; and our proposal, using incremental distance calculations. There is a problem with the line search approach, because the set of points it identifies may include spurious corner-points. The proposed algorithm, however, maintains the quality of, and is three times faster, than the classical algorithm.

An efficient image compression algorithm without distortion

Huang, L. and A. Bijaoui. An efficient image compression algorithm without distortion, Pattern Recognition Letters 12 (1991) 69–72. We present in this paper an efficient image compression algorithm without distortion. It is based on a special iterative central prediction method with a hierarchic 4-bit code. By applying this method to some hydrodynamic simulation images, high compression ratios have been obtained.

Minimizing the cost of errors with a Markov random field

Sher, D.B., Minimizing the cost of errors with a Markov random field, Pattern Recognition Letters 12 (1991) 85–89. Marroquin (1985) and Dinton et al. (1988) argue that the MPM estimator is superior to the MAP estimator for vision algorithms. We translate a Markov random field into another Markov random field whose MAP estimate is the MPM estimate of the original field. Our technique uses global optimization algorithms such as simulated annealing to compute a MPM estimate for a Markov random field.

Recognition and positioning of partially occluded 3-D objects

Ray, K.S. and D. Dutta Majumder, Recognition and positioning of partially occluded 3-D objects, Pattern Recognition Letters 12 (1991) 93–108. The task of recognizing and positioning the partially occluded three-dimensional (3-D) rigid objects of a given scene is considered. The surfaces of 3-D objects may be planar or curved. The 3-D surface informations are captured through range data (depth map). For recognition we use the principal curvatures, mean curvature and Gaussian curvature as the local descriptions of the surfaces. These curvatures are the local invariant features of the surfaces. A computer vision scheme, based upon the matching between the local features of the 3-D objects in a scene and those of the models which are considered as knowledge data base, is described. Finally, the hypothesis generation and verification scheme is considered for best possible recognition.

An efficient transportation algorithm for automatic chromosome karyotyping

Tso, M., P. Kleinschmidt, I. Mitterreiter and J. Graham, An efficient transportation algorithm for automatic chromosome karyotyping, Pattern Recognition Letters 12 (1991) 117–126. We have implemented an algorithm for the special case of the transportation problem with unit demands to assist in the automatic classification (karyotyping) of human chromosomes by image analysis. Use of the algorithm permits prior knowledge of the number of chromosomes of a certain type in a normal human cell to constrain the classification. A study involving the classification of three large datasets is described and a comparison is made of the maximum likelihood and Bayesian approaches.

Parallel algorithms for testing if a point is inside a closed curve

Chen, L.T. and L.S. Davis, Parallel algorithms for testing if a point is inside a closed curve, Pattern Recognition Letters 12 (1991) 73–77. This paper describes two parallel algorithms to test whether a point is inside a digitized closed curve. The first algorithm assumes the curve is stored in a linear array of processors. The second algorithm assumes the digital image containing the curve is stored in a 2-D array of processors and that a clockwise (or counterclockwise) trace of the curve is available (i.e., each point on the curve has a pointer to the next clockwise point). Both algorithms assume that scan instructions are available on the processor array. The algorithms take constant numbers of scan instructions.

Fold principal axis—a new tool for defining the orientations of rotationally symmetric shapes

Chou, S.-L., J.-C. Lin and W.-H. Tsai, Fold principal axis—a new tool for defining the orientations of rotationally symmetric shapes, Pattern Recognition Letters 12 (1991) 109–115. A new type of principal axis, namely, fold principal axis, is introduced in this paper, which can be used to define the orientations of rotationally symmetric shapes. A given shape is first transformed into a new shape on which the well-defined traditional principal axis is detected. The traditional principal axis is then transformed back to obtain the proposed fold principal axes of the given shape. The properties of such axes, including the uniqueness, existence, and invariance under rotation, translation, and scaling, are investigated. Some illustrative examples are also given.

Comments on grey-level thresholding of images using a correlation criterion

Brink, A.D., Comments on grey-level thresholding of images using a correlation criterion, Pattern Recognition Letters 12 (1991) 91–92. I refer to the comments of Cseke and Fazekas concerning the thresholding algorithms proposed by Otsu and Brink, respectively. Contrary to the claim made by Cseke and Fazekas it is shown that these algorithms are not generally identical, this case arising only when the black/white levels used for the bilevel image are the below- and above-threshold means of the original image.