This study investigated how three faculty members and an external lecturer selected training content and implemented teaching methods for faculty development (FD) to ensure learning opportunities for faculty members at science universities. Qualitative analysis revealed the following three points: (1) the lack of a clear image of the target participants causes discrepancies between participant needs and the FD seminar planners; (2) the faculty developer (FDer) is apt to “chase after two hares,” when seeking student-centered teaching style and instructors’ language skills improvement; and (3) FD seminars should be focused on a specific topic to be more attractive to the target participants.

Gaussian smoothing and its first and second differentials are very important for image processing and computer vision. We already have methods based on the sliding Fourier transform (SFT) in order to calculate Gaussian smoothing efficiently. However, errors in floating point calculation accumulate quickly when we use the single-precision floating-point format. Also, the resultant discontinuities at the edges of SFT intervals generate false extreme points easily. To resolve the abovementioned problems, we propose the attenuated SFT (ASFT) by introducing a decay factor to the original SFT together with a new criterion for determining coefficients to effectively approximate Gaussian function and its differentials. Extensive experiments demonstrate a decisive superiority of the proposed ASFT over the original SFT.

The guided filter and its subsequent derivatives have been widely employed in many image processing and computer vision applications primarily brought about by their low complexity and good edge-preservation properties. Despite this success, the different variants of the guided filter are unable to handle more aggressive filtering strengths leading to the manifestation of "detail halos". At the same time, these existing filters perform poorly when the input and guide images have structural inconsistencies. In this paper, we demonstrate that these limitations are due to the guided filter operating as a variable-strength locally-isotropic filter that, in effect, acts as a weak anisotropic filter on the image. Our analysis shows that this behaviour stems from the use of unweighted averaging in the final steps of guided filter variants including the adaptive guided filter (AGF), weighted guided image filter (WGIF), and gradient-domain guided image filter (GGIF). We propose a novel filter, the Anisotropic Guided Filter (AnisGF), that utilises weighted averaging to achieve maximum diffusion while preserving strong edges in the image. The proposed weights are optimised based on the local neighbourhood variances to achieve strong anisotropic filtering while preserving the low computational cost of the original guided filter. Synthetic tests show that the proposed method addresses the presence of detail halos and the handling of inconsistent structures found in previous variants of the guided filter. Furthermore, experiments in scale-aware filtering, detail enhancement, texture removal, and chroma upsampling demonstrate the improvements brought about by the technique.

Edge-preserving filtering is vital in many image processing and computer vision tasks. However, existing techniques, such as the bilateral filter and the guided filter, have limitations when dealing with stronger filtering strengths, resulting in the suppression of image structures. In this work, we discuss an alternative structure-preserving image filter (SPIF) that operates on multiple detail densities (i.e. scales) so as to overcome the trade-off between filtering strength and edge-preservation. We show, using several experiments, that the proposed filter is capable of performing various tasks, specifically detail enhancement and image abstraction, while maintaining low computational times. The parameterization of the proposed filter provides a high degree of flexibility, allowing it to perform well for the aforementioned tasks, and potentially, in many other tasks.

GAT (Global Affine Transformation) and GPT (Global Projection Transformation) correlation matchings were successively proposed by Wakahara and Yamashita which use affine transformation (AT) and 2D projection transformation (PT), respectively, to maximize the normalized cross-correlation value between a template and a GAT/GPT-superimposed input image. In theory, to maximize the degree of matching via normalized cross-correlation, the L2 norm of both images should be normalized. However, the criteria of conventional GAT/GPT correlation techniques did not take account of the conservation of the L2 norm. This research solves the above-mentioned problem which might impair the matching ability of GAT/GPT correlations. In particular, we focus on the enhanced GPT correlation recently proposed by Wakahara et al. to calculate optimal PT parameters simultaneously. Then, we propose an improved criterion satisfying the requirement of the L2 norm conservation together with the enhanced GPT correlation algorithm with norm normalization. Experimental results using artificially deformed images of chess board and handwritten digits from MNIST database by 2D projection transformation show that the proposed method clearly outperforms those conventional GPT correlation techniques with or without norm normalization in terms of both convergence speed and matching ability.

In this work, we tackle the problem of coloring black-and-white images, which is image colorization. Existing image colorization algorithms can be categorized into two types: scribble-based colorization algorithms and example-based colorization algorithms. Differently, we propose a hybrid scheme that combines the advantages of both categories. Given the grayscale image to be colorized and a few color scribbles (or scattered color labels) as input, the proposed method manages to colorize the grayscale image with high quality. Similar to the mechanisms in example-based colorization methods, our algorithm firstly propagates chrominance information based on the assumption that similar image patches should have similar colors. Therefore colors of some pixels can be transferred from similar patches with known colors. After that, we apply scribble-based colorization algorithm to fully colorize the grayscale image, with different confidences assigned onto the transferred color labels. Experimental results show that, the proposed method effectively utilizes the known chrominance, and provides pleasant colorizations with very few user interventions.

This paper describes a new technique of 2D projection transformation invariant template matching, GPT (Global Projection Transformation) correlation, as a natural extension of our earlier work on the affine-invariant GAT (Global Affine Transformation) correlation method. The key ideas are threefold. First, we show that arbitrary 2D projection transformation (PT) can be decomposed into a product of affine transformation (AT) and partial projection transformation (PPT). Second, we propose an efficient computational model for determining sub-optimal components of AT and PPT separately that maximize a normalized cross-correlation value between an either AT- or PPT-superimposed input image and a template by solving linearized simultaneous equations. Third, we obtain optimal components of combined AT and PPT, i.e. PT, that maximize a normalized cross-correlation value between a PT-superimposed input image and a template via the successive iteration method. The proposed technique has the time complexity of O(n2), where n equals the number of pixels. Experiments using templates and their artificially distorted images as input images show that the proposed method is far superior to the GAT correlation method in 2D projection transformation tolerance, and, also, has a high tolerance for noise.

The global affine transformation (GAT) correlation method proposed by Wakahara et al. is a pattern matching method that can compensate for affine transformation embedded in an input pattern. The GAT correlation method demonstrated a high performance in character recognition and object matching. For example, it outperformed the well-known tangent distance (TD) method in character recognition experiments made on IPTP handwritten numeral database. The purpose of this paper is threefold. First, we introduce a new matching measure called the nearest neighbor distance of equi-gradient direction (NNDEGD) in cooperation with the GAT correlation method. The NNDEGD is just the window parameter of the Gaussian function used in the GAT correlation method, which is equal to the average minimum distance between a point in one image and another point in the other image with the same gradient direction. We propose to use this value as a new matching measure. Secondly, we extend the GAT correlation method so as to handle the change of stroke width besides the affine transformation. Finally, we apply the original and extended versions of the GAT correlation method to fc-NN classification experiments using the MNIST database. These experiments are carried out efficiently for the first time because we have substantially reduced the computational complexity and memory load involved in the original GAT correlation method. We successfully show that the enhanced GAT correlation method has achieved the lowest error rate of 0.49% compared with competing distortion-tolerant template matching techniques.

Fisher discriminant analysis (FDA) is a very famous analysis method for classification. However, FDA does give an optimal projection only for Gaussian distributions with equal covariance matrices. In other words, FDA is not optimal in the case of heteroscedastic Gaussian distributions. In this paper, we propose a novel criterion for FDA including a correction term based on the Bhattacharyya distance which is closely related to classification rate. Furthermore, the Chernoff distance based criterion and its kernelized version are proposed as its extension. These proposed criteria have three strong points. The first one is that the correction term based on the Bhattacharyya distance can deal with heteroscedastic Gaussian distributions. The second one is that the correction term based on the Chernoff distance can handle easily the difference in the number of class samples. The third point is that their kernel extensions are easily implemented to be applied to non-Gaussian distributions. As a result, the proposed method is applicable in a wide variety of classification problems. Experimental results using toy simulations and nine kinds of UCI real-world datasets show marked advantages of the proposed method over the conventional FDA.

In this paper, we propose classifiers derived from quadratically constrained maximum a posteriori (QCMAP) estimation. The QCMAP consists of the maximization of the expectation of a cost function, which is derived from the maximum a posteriori probability and a quadratic constraint. This criterion is highly general since its forms include least squares regressions and a support vector machine. Furthermore, the criterion provides a novel classifier, the "Gaussian QCMAP." The QCMAP procedure still has large theoretical interest and its full extensibility has yet to be explored. In this paper, we propose using the mixture of Gaussian distributions as the QCMAP weight function. The mixture of Gaussian distributions has wide-ranging applicability, and encompasses forms, such as a normal distribution model and a kernel density model. We propose four types of mixture of Gaussian functions for QCMAP classifiers, and conduct experiments to demonstrate their advantages.

This paper addresses the problem of reinforcing the ability of k-NN classification of handwritten characters via distortion-tolerant template matching techniques with a limited quantity of data. We make a comparison of three kinds of matching techniques: the conventional simple correlation, the tangent distance, and the GAT (Global Affine Transformation) correlation. The k-NN classification method is straightforward and powerful, however, is very time-consuming. Hence, to reduce the computational cost of matching in k-NN classification we propose to accelerate the GAT correlation technique by reformulating its computational model and adopting efficient lookup tables. Recognition experiments made on the handwritten numerical database IPTP CDROM1B show that matching techniques of the simple correlation, the tangent distance, and the accelerated GAT correlation achieve recognition rates of 97.07%, 97.50%, and 98.70%, respectively. Also, the computation time ratios of the tangent distance and the accelerated GAT correlation to the simple correlation are 26.3 and 36.5 to 1.0, respectively.

Maximum a posteriori (MAP)-based kernel classification trained by linear programming (MAPLP) has previously been proposed as a new approach to MAP-based classification. As opposed to conventional MAP-based approaches, MAPLP does not directly estimate a posteriori probabilities for classification, but instead introduces its surrogate function to an objective function that behaves similarly to a MAP-based classifier. In this paper, two extensions of MAPLP are proposed. In the first extension we add an adjustment parameter to one of the MAPLP constraints, and in the second extension we apply a difference-type objective function. Moreover, binary classification experiments are performed to assess the performance of the MAPLP extensions. The results are compared with those of the standard MAPLP as well as those of other state-of-the-art classification methods, such as support vector machines and kernel Fisher discriminant analysis. The experimental results not only demonstrate that the MAPLP extensions provide improved performance over standard MAPLP for a number of well-known datasets, they also perform promisingly against other classification methods.

This paper proposes a new, affine-invariant image matching technique via accelerated KL (Kullback-Leibler) divergence minimization. First, we represent an image as a probability distribution by setting the sum of pixel values at one. Second, we introduce affine parameters into either of the two images' probability distributions using the Gaussian kernel density estimation. Finally, we determine optimal affine parameters that minimize KL divergence via an iterative method. In particular, without using such conventional nonlinear optimization techniques as the Levenberg-Marquardt method we devise an accelerated iterative method adapted to the KL divergence minimization problem through effective linear approximation. Recognition experiments using the handwritten numeral database IPTP CDROM1B show that the proposed method achieves a much higher recognition rate of 91.5% at suppressed computational cost than that of 83.7% obtained by a simple image matching method based on a normal KL divergence.

This paper addresses the problem of improving the accuracy of character recognition with a limited quantity of data. The key ideas are twofold. One is distortion-tolerant template matching via hierarchical global/partial affine transformation (GAT/PAT) correlation to absorb both linear and nonlinear distortions in a parametric manner. The other is use of multiple templates per category obtained by k-means clustering in a gradient feature space for dealing with topological distortion. Recognition experiments using the handwritten numerical database IPTP CDROM1B show that the proposed method achieves a much higher recognition rate of 97.9% than that of 85.8% obtained by the conventional, simple correlation matching with a single template per category. Furthermore, comparative experiments show that the k-NN classification using the tangent distance and the GAT correlation technique achieves recognition rates of 97.5% and 98.7%, respectively.

The discriminant function of a subspace method is provided by using correlation matrices that reflect the averaged feature of a category. As a result, it will not work well on unknown input patterns that are far from the average. To address this problem, we propose two kinds of weighted correlation matrices for subspace methods. The globally weighted correlation matrix (GWCM) attaches importance to training patterns that are far from the average. Then, it can reflect the distribution of patterns around the category boundary more precisely. The computational cost of a subspace method using GWCMs is almost the same as that using ordinary correlation matrices. The locally weighted correlation matrix (LWCM) attaches importance to training patterns that are near to an input pattern to be classified. Then, it can reflect the distribution of training patterns around the input pattern in more detail. The computational cost of a subspace method with LWCM at the recognition stage does not depend on the number of training patterns, while those of the conventional adaptive local and the nonlinear subspace methods do. We show the advantages of the proposed methods by experiments made on the MNIST database of handwritten digits.

We propose a novel efficient method of blind signal extraction from multi-sensor networks when each observed signal consists of one global signal and local uncorrelated signals. Most of existing blind signal separation and extraction methods such as independent component analysis have constraints such as statistical independence, non-Gaussianity, and underdetermination, and they are not suitable for global signal extraction problem from noisy observations. We developed an estimation algorithm based on alternating iteration and the smart weighted averaging. The proposed method does not have strong assumptions such as independence or non-Gaussianity. Experimental results using a musical signal and a real electroencephalogram demonstrate the advantage of the proposed method.

The Wiener filter (WF) is widely used for inverse problems. From an observed signal, it provides the best estimated signal with respect to the squared error averaged over the original and the observed signals among linear operators. The kernel WF (KWF), extended directly from WF, has a problem that an additive noise has to be handled by samples. Since the computational complexity of kernel methods depends on the number of samples, a huge computational cost is necessary for the case. By using the first-order approximation of kernel functions, we realize KWF that can handle such a noise not by samples but as a random variable. We also propose the error estimation method for kernel filters by using the approximations. In order to show the advantages of the proposed methods, we conducted the experiments to denoise images and estimate errors. We also apply KWF to classification since KWF can provide an approximated result of the maximum a posteriori classifier that provides the best recognition accuracy. The noise term in the criterion can be used for the classification in the presence of noise or a new regularization to suppress changes in the input space, whereas the ordinary regularization for the kernel method suppresses changes in the feature space. In order to show the advantages of the proposed methods, we conducted experiments of binary and multiclass classifications and classification in the presence of noise.

Blind source extraction is one of the most important problems for multi-sensor networks. We propose a blind source extraction and deconvolution method in the presence of noise. We use MA-model for the signal generation model, and the convolutive observation model. The parameter of MA-model and the observations are obtained from an alternating least square (ALS) algorithm. The reconstruction is done by an spatiotemporal inverse filter such that it minimizes the Euclidean distance between the original signal and the reconstruction signal. Experimental results demonstrate advantages of the proposed method.

The Mahalanobis metric was proposed by extending the Mahalanobis distance to provide a probabilistic distance for a non-normal distribution. The Mahalanobis metric equation is a nonlinear second order differential equation derived from the equation of geometrically local isotropic independence, which is proposed to define normal distributions in a manifold. In this paper we provide experimental results of calculating the Mahalanobis metric by the Newton-Raphson method. We add error to the original probability density function and calculate the Mahalanobis metric to investigate the effect of the error in a probability density function to the solution.

We propose a new classification method based on a kernel technique called suppressed kernel sample space projection classifier (SKSP), which is extended from kernel sample space projection classifier (KSP). In kernel methods, samples are classified after they are mapped from an input space to a high-dimensional space called a feature space. The space that is spanned by samples of a class in a feature space is defined as a kernel sample space. In KSP, an unknown input vector is classified to the class of which projection norm onto the kernel sample space is maximized. KSP can be interpreted as a special type of kernel principal component analysis (KPCA). KPCA is also used in classification problems. However, KSP has more useful properties compared with KPCA, and its accuracy is almost the same as or better than that of KPCA classifier. Since KSP is a single-class classifier, it uses only self-class samples for learning. Thus, for a multiclass classification problem, even though there are very many classes, the computational cost does not change for each class. However, we expect that more useful features can be obtained for classification if samples from other classes are used. By extending KSP to SKSP, the effects of other classes are suppressed, and useful features can be extracted with an oblique projection. Experiments on two-class classification problems, indicate that SKSP shows high accuracy in many classification problems.

A class of lapped transforms for image coding, which are characterized by variable-length synthesis filters, is introduced. In this class, the synthesis filter bank (FB) is first defined with an arbitrary combination of finite impulse response synthesis filters of perfect reconstruction FBs. An analysis FB is then obtained using direct matrix inversion or iterative implementation of Neumann series expansion. Moreover, to improve compression, we introduce a unitary transform that follows the analysis FB. This class enables a greater freedom of design than previously presented variable-length lapped transforms. We illustrate several design examples and present experimental results for image coding, which indicate that the proposed transforms are promising and comparable with conventional subband transforms including wavelets.

The theory and design for a novel class of a variable-length lapped transform are presented. In this class, the synthesis filter bank is firstly defined with arbitrary combination of FIR filters, giving much freedom of design compared to previously presented variable-length lapped transforms. The analysis filter bank is then obtained as the inverse of the synthesis bank by the Neumann series expansion to achieve the perfect reconstruction. Moreover, by introducing a unitary transform, we improve the compression performance of the analysis bank. We provide several design examples and experimental results of image coding, which show that the proposed transform is promising and comparable with conventional subband transforms including wavelets.

A novel time-varying subband transform with projection-based reconstruction and its application in image coding are introduced. At each time instance (block), the transform to be applied is chosen from a given set of orthogonal subband transforms according to a certain criterion. The reconstruction from the transformed coefficients is possible by applying the theory of convex projections (POCS). We introduce convex sets for perfect reconstruction from the transformed vector. We further propose other convex sets which assure smoothness of plane regions in the decoded image. Experimental results show that the use of multiple subband transforms leads to the improvement of coding performance compared to the use of single transform.

We investigate a special class of N-channel oversampled linear-phase perfect reconstruction filter banks with a decimation factor M smaller than N. We deal with systems in which all analysis and synthesis filters have the same FIR length and share the same center of symmetry. We provide the general lattice factorization of a polyphase matrix of a particular class of these oversampled filter banks. The lattice structure is based on the singular value decomposition for non-square matrices. The resulting lattice structure is able to provide fast implementation and allows us to determine the filter coefficients by solving an unconstrained optimization problem. We show that the present systems with the lattice structure cover a wide range of linear-phase perfect reconstruction filter banks. We also show several design examples.

An oversampled linear phase (LP) perfect reconstruction (PR) filter bank (FB) with a lattice structure involving a noise robust building block is presented. The structure is based on the generalized lapped pseudo-biorthogonal transform (GLPBT), which represents a large class of lapped transforms. Given an analysis bank of the GLPBT and a correlation matrix of noise, the proposed synthesis bank reduces the noise component and attains PR. Numerical results show that this class of filter bank is effective in transmission of oversampled data over noisy channels.

This paper presents a new framework for a biorthogonal lapped transform that consists of long and short basis functions called the VLLBT. It is shown that when the biorthogonal long basis functions of the VLLBT are given, the optimal short basis functions in the energy compaction sense are derived by solving an eigenvalue problem without iterative searching techniques. We also provide design and image coding examples of the VLLBT. The resulting VLLBT attains a high coding gain compared to other lapped transforms. Moreover, experimental results show that the proposed VLLBT is superior to other conventional transforms in terms of PSNR at a high compression ratio. Furthermore, it significantly reduces the annoying blocking artifacts.

A lapped biorthogonal transform adapted for local directionality of an image is proposed. The proposed transform consists of overlapping basis functions and non-overlapping basis functions, where the basis functions' centers of symmetry are aligned. Because of the alignment, we can treat the image boundaries without special processing. In order to construct non-overlapping basis functions, we also introduce a transform that provides the optimal approximation of an original signal in a given subspace. In image coding, an image block is selectively transformed by one of the orientation adaptive transforms. Experimental results show that the use of adaptation improves the quality of decoded images.

A lapped orthogonal transform adapted for local directionality of an image is proposed and its application in adaptive image coding is presented. First, we address the problem of constructing the lapped orthogonal transform with short and long functions such that the long functions are twice the length of the short functions, and derive the transform with the optimal short functions in the energy compaction sense. Moreover, using the optimal solution, we propose the concept of an orientation adaptive lapped orthogonal transform (OALOT) and give its design method. Experimental results show that in addition to the property of the LOT which reduces blocking artifacts, the OALOT possesses the property of preserving strong edges and lines.

In this paper, the theory and the design of a new class of orthogonal transforms are presented. The novel transform is derived from a correlation matrix in which an arbitrary orthonormal system is embedded. By embedding an orthonormal system designed empirically, we obtain the transform that can not only represent intuitive features but also possesses statistical property like the KLT. Our main motivation is the application in block-based adaptive transforms coding. We show a design example of the transform, which adapts orientational features such as edges and lines. Using this transform, we perform orientation adaptive coding. In experimental results, it is shown that image coding using the transform is effective in rate-distortion criterion and subjective quality.

CLAFIC (class-featuring information compression) is a well-known class feature extraction method. By using Karhunen-Loeve transform (KLT) for patterns in a category, class features for the category are extracted. However, such a class feature may not be suitable for classification, if it is also contained in other categories. Suitable class features for classification have to be contained in a category but not in the other categories. In order to solve this problem, we propose the relative Karhunen-Loeve transform method (RKLTM) for class feature extraction. We show the advantages of RKLTM over CLAFIC by the experiments on handwritten numeral recognition.

We define the weighted Karhunen-Loeve transform (WKLT) and provide its general form. For use in image compression we propose a new method to reduce the blocking effects by using overlapping blocks and the WKLT. Although by decreasing the weight in the overlapping areas, the error in these areas will increase, the error in the non-overlapping areas will decrease. The error increased in the overlapping areas can be reduced by averaging over blocks. Thus, the error on each pixel is almost the same, and the redundancy for the overlapping areas is reduced.

The Karhunen-Loeve (K-L) subspace is a subspace which provides the best approximation for a stochastic signal under the condition that its dimension is fixed. The K-L subspace, however, does not consider a noise in communication and a noise and patterns in other categories in pattern recognition. Therefore, its noise suppression is not sufficient in communication. It gives a wrong recognition result for similar patterns but belong to different categories. In order to solve this problem, we propose relative K-L operators. The advantage of the relative K-L operators is illustrated by using a simulation.