In statistical pattern recognition, it is important to estimate the distribution of patterns precisely to achieve high recognition accuracy. In general, precise estimation of the parameters of the distribution requires a great number of sample patterns, especially when the feature vector obtained from the pattern is high-dimensional. For some pattern recognition problems, such as face recognition or character recognition, very high-dimensional feature vectors are necessary and there are always not enough sample patterns for estimating the parameters. In this paper, we focus on estimating the distribution of high-dimensional feature vectors with small number of sample patterns. First, we define a function, called simplified quadratic discriminant function (SQDF). SQDF can be estimated with small number of sample patterns and approximates the quadratic discriminant function (QDF). SQDF has fewer parameters and requires less computational time than QDF. The effectiveness of SQDF is confirmed by three types of experiments. Next, as an application of SQDF, we propose an algorithm for estimating the parameters of the normal mixture. The proposed algorithm is applied to face recognition and character recognition problems which require high-dimensional feature vectors.

Precise estimation of data distribution with a small number of sample patterns is an important and challenging problem in the field of statistical pattern recognition. In this paper, we propose a novel method for estimating multimodal data distribution based on the Gaussian mixture model. In the proposed method, multiple random vectors are generated after classifying the elements of the feature vector into subsets so that there is no correlation between any pair of subsets. The Gaussian mixture model for each subset is then constructed independently. As a result, the constructed model is represented as the product of the Gaussian mixture models of marginal distributions. To make the classification of the elements effective, a graph cut technique is used for rearranging the elements of the feature vectors to gather elements with a high correlation into the same subset. The proposed method is applied to a character recognition problem that requires high-dimensional feature vectors. Experiments with a public handwritten digit database show that the proposed method improves the accuracy of classification. In addition, the effect of classifying the elements of the feature vectors is shown by visualizing the distribution.

Reducing the time complexity of character matching is critical to the development of efficient Japanese Optical Character Recognition (OCR) systems. To shorten the processing time, recognition is usually split into separate pre-classification and precise recognition stages. For high overall recognition performance, the pre-classification stage must both have very high classification accuracy and return only a small number of putative character categories for further processing. Furthermore, for any practical system, the speed of the pre-classification stage is also critical. The associative matching (AM) method has often been used for fast pre-classification because of its use of a hash table and reliance on just logical bit operations to select categories, both of which make it highly efficient. However, a certain level of redundancy exists in the hash table because it is constructed using only the minimum and maximum values of the data on each axis and therefore does not take account of the distribution of the data. We propose a novel method based on the AM method that satisfies the performance criteria described above but in a fraction of the time by modifying the hash table to reduce the range of each category of training characters. Furthermore, we show that our approach outperforms pre-classification by VQ clustering, ANN, LSH and AM in terms of classification accuracy, reducing the number of candidate categories and total processing time across an evaluation test set comprising 116,528 Japanese character images.

In speech recognition, confidence annotation adopts a single confidence feature or a combination of different features for classification. These confidence features are always extracted from decoding information. However, it is proved that about 30% of knowledge of human speech understanding is mainly derived from high-level information. Thus, how to extract a high-level confidence feature statistically independent of decoding information is worth researching in speech recognition. In this paper, a novel confidence feature extraction algorithm based on latent topic similarity is proposed. Each word topic distribution and context topic distribution in one recognition result is firstly obtained using the latent Dirichlet allocation (LDA) topic model, and then, the proposed word confidence feature is extracted by determining the similarities between these two topic distributions. The experiments show that the proposed feature increases the number of information sources of confidence features with a good information complementary effect and can effectively improve the performance of confidence annotation combined with confidence features from decoding information.