International Journal of Pattern Recognition and Artificial Intelligence

We present an approach to automatically extract a pertinent subset of soft output classifiers, and to aggregate them into a global decision rule using the Choquet integral. This approach relies on two key points. The first is a learning algorithm that uses a measure of the confusion between the categories to be recognized. The second is a selection scheme that discards weak or redundant decision rules, keeping only the most relevant subset. An experimental study, based on real world data, is then described. It analyzes the improvements achieved by these points first when used independently, then when combined together.

License plate recognition systems have been used extensively for many applications including parking lot management, tollgate monitoring, and for the investigation of stolen vehicles. Most researches focus on static systems, which require a clear and level image to be taken of the license plate. However, the acquisition of images that can be successfully analyzed relies on both the location and movement of the target vehicle and the clarity of the environment. Moreover, only few studies have addressed the problems associated with instant car image processing. In view of these problems, a real-time license plate recognition system is proposed that recognizes the video frames taken from existing surveillance cameras. The proposed system finds the location of the license plate using projection analysis, and the characters are identified using a back propagation neural network. The strategy achieves a recognition rate of 85.8% and almost 100% after the neural network has been retrained using the erroneously recognized characters, respectively.

In this paper, we propose a new method, named Greedy Equivalence Search-Expectation Maximization (GES-EM), for learning Bayesian networks from incomplete data. Our method extends the recently proposed Greedy Equivalence Search (GES) algorithm¹⁰ to deal with incomplete data. For the quality evaluation of learned networks, we make use of the expected Bayesian Information Criterion (BIC) scoring function. In addition, we propose a new structural evaluation criterion. This so-called SEC criterion is more suitable than existing structural evaluation criteria, since it is based on the comparison of learned networks to the generating ones through Completed Partially Directed Acyclic Graphs (CPDAGs). Experimental results show that GES-EM algorithm yields more accurate structures than the standard Alternating Model Selection-Expectation Maximization (AMS-EM) algorithm.¹⁵.

We present a new method for simultaneously determining three-dimensional (3D) motion and structure of a nonrigid object from its uncalibrated two-dimensional (2D) data with Gaussian or non-Gaussian distributions. A nonrigid motion can be treated as a combination of a rigid component and a nonrigid deformation. To reduce the high dimensionality of the deformable structure or shape, we estimate the probability distribution function (PDF) of the structure through random sampling, integrating an established probabilistic model. The fitting between the observations and the estimated 3D structure will be evaluated using the pooled variance estimator. The recovered structure is only available when the 2D feature points have been properly corresponded over two image frames. Applications of the proposed method to both synthetic and real image sequences are demonstrated with promising results.

Many techniques and algorithms for automatic text categorization had been devised and proposed in the literature. However, there is still much space for researchers in this area to improve existing algorithms or come up with new techniques for text categorization (TC). Polynomial Networks (PNs) were never used before in TC. This can be attributed to the huge datasets used in TC, as well as the technique itself which has high computational demands. In this paper, we investigate and propose using PNs in TC. The proposed PN classifier has achieved a competitive classification performance in our experiments. More importantly, this high performance is achieved in one shot training (noniteratively) and using just 0.25%–0.5% of the corpora features. Experiments are conducted on the two benchmark datasets in TC: Reuters-21578 and the 20 Newsgroups. Five well-known classifiers are experimented on the same data and feature subsets: the state-of-the-art Support Vector Machines (SVM), Logistic Regression (LR), the k-nearest-neighbor (kNN), Naive Bayes (NB), and the Radial Basis Function (RBF) networks.

In order to improve the performance in pattern classification, we utilize multiple classifiers and combine their individual decisions to make a final decision. In this paper, we present the combination using Bayesian method and compare minimum errors. This method requires the posteriori probabilities from all classifiers, which may be difficult to calculate in real world because tremendous amounts of training samples are needed. Alternatively, a confusion matrix is developed for approximation. We also use different combining rules for comparisons and apply them to handwritten digit recognition.

Image Quality Measure (IQM) is used to automatically measure the degree of image artifacts such as blocking, ringing and blurring effects. It is calculated traditionally in the image spatial domain. In this paper, we present a new method of transforming an image into a low-dimensional domain based on random projection, so we can efficiently obtain the compatible IQM. From the transformed domain, we can calculate the Peak Signal-to-Noise Ratio (PSNR) and apply fuzzy logic to generate a Low-Dimensional Quality Index (LDQI). Experimental results show that the LDQI can approximate the IQM in the image spatial domain. We observe that the LDQI is suited for measuring the compression blur due to its relatively low distortion. The relative error is about 0.15 as the compression blur increases.

In this article, we present a new fingerprint classification algorithm. Singular points are first extracted from enhanced fingerprint direction images with a resolution of 2 × 2 pixels by the modified SEA algorithm. Based on the number of singular points, fingerprints are categorized into types of "arch", "whorl", and "solitary". Solitary fingerprints are properly rotated and then further processed to generate direction patterns that lead to establishment of individual direction template. Direction constraints are formed and derived from pattern descriptors by their structural layout. Decision rules are then established and pattern templates are classified into three more types: "right loop", "left loop", and "tented arch". NIST-4 database was used for an experimental test, and our classification accuracy was 91.62% with 1.55% rejection on the five-class system (94.38% on the four-class system), which is the best result on the five-class system to-date. An additional experiment on NIST-14 database reports 89.15% accuracy with 3.07% rejection.