Search

An approach based on stroke orientation and asymmetric distribution model about feature parameter was proposed, which incorporates structural feature into statistical strategy. An improvement to the algorithm for extracting fork points from skeleton images defends the reliability of stroke extraction. The feature vector for statistical recognition is extracted directly from stroke structure, and the asymmetric distribution model is applied to compute distances. The experimental results indicate that the proposed system is effective to handwritten Chinese character recognition. [ABSTRACT FROM AUTHOR]

A novel diversity-sampling based Gaussian kernel density estimation (KDE) model was proposed for the representation of multimodal background. Choosing those samples that have diversiform gray-levels in training sequence, a nonparametric model was built for modeling the scene background. According to the related gray-level, the different weights are given to the different samples in kernel density estimation. This avoids the repetition computation using the total samples, and makes KDE very effective. The experimental results show the good detection performance in the traffic surveillance system. [ABSTRACT FROM AUTHOR]

The purpose of trace reconstruction is to recover information about handwriting sequence from static images of characters, which helps incorporate online methods into offline applications and unify the recognition strategies of single character and character sequence. A stroke segment-based algorithm was proposed, which is equal to the problem of ordering the stroke segments in nature. Structural graph of stroke segments is extracted from skeleton images, by which the relational graph is created. Trace reconstruction is realized as a globally optimal problem, and the handwriting trace is considered as the path with the totally minimal orientation variance, which can be resolved by searching a Hamiltonian path with minimal cost. The cases of trace reconstruction were analyzed, and the accuracy reaches 93.5% on 200 images. The experimental results indicate the proposed approach is effective to the reconstruction of handwriting traces of handwritten numerals. [ABSTRACT FROM AUTHOR]

A simple but effective algorithm which is based on minimizing the maximum discrepancy clustering was presented. Two new metrics of "distance" between any two triangles is developed. By the new distance metrics, the clustering method is easily extended to divide the input mesh into several connected regions. Furthermore, a post-processing algorithm, constrained boundary straightening, was proposed to regularize the shapes of partitioned regions. The experiments show that this two-step solution for mesh segmentation performs well in both region planarity and region shape. [ABSTRACT FROM AUTHOR]

This paper presented an integrated method for unlabeled data assisted active learning in order to solve insufficient labeled training data in supervised learning. The proposed approach uses naive Bayes with expectation maximization (EM) to learn an initial query concept based on the labeled and formerly unlabeled images. Then the active learning algorithm selects the most useful images in database to query the user for labeling, and the query concept is updated iteratively according to user's feedbacks. Various experiments demonstrate that the proposed approach can improve the efficiency and accuracy of the learner and can grasp user's query concept quickly. [ABSTRACT FROM AUTHOR]

When the texture of an iris image is normalized, there is always fatal interference: center deviation of the iris' internal and external circles. This interference has to be corrected in order to ensure an accurate normalization, an essential step of image preprocessing. The reasons of deviation were discussed, based on ophthalmology and image gathering, and then the first and second deviation calibration algorithm were described. The experiments on 400 images show that these two algorithms are of much better normalization results than the traditional one and higher confidence level. [ABSTRACT FROM AUTHOR]

An approach for lossless compression of still images based on fuzzy-combined adaptive prediction was proposed. In the phase of prediction preparation, a set of optimal autoregressive predictors is calculated adaptively based on a predetermined threshold. Then the predictors are trained by a neural network called FOSART. The predictors attained from procedure of training are incorporated in sense of fuzzy logic to implement the procedure of prediction. In the entropy coding phase, the context-based conditional adaptive arithmetic encoding is adopted in the proposed approach. The experiments demonstrate the characteristics make the approach achieve good tradeoff between computational complexity and efficiency of prediction and it has a good performance for lossless compression. [ABSTRACT FROM AUTHOR]

Segmentation of off-line handwritten Chinese characters is the premise of recognition. It is most difficult to segment connected characters. A novel algorithm based on stroke analysis and background thinning was proposed to segment connected handwritten Chinese characters. The feature points, viz. end points, fork points and corner points are detected in the thinned image of connected characters. The segments between the feature points are considered as substrokes and are extracted. The information of the length of substrokes, the topological relations and projection is employed to locate connected points. By thinning the background, suitable separation path is determined. The experimental results show that satisfactory performance is achieved by the presented method for segmentation of connected handwritten Chinese characters. [ABSTRACT FROM AUTHOR]

As a preprocessing, the text area location plays an important role in the character recognition. This paper presented a morphological method to locate text areas for character images. By some morphological operations, the proposed method can locate text area from a complex image quickly and accurately. An optimum method was proposed which can be used in the real time circumstance. The approach was tested in a real application to locate PCB characters and got a good result. [ABSTRACT FROM AUTHOR]

Multi agent system is a distributed processing system which needs an opened and dynamic knowledge collaborative architecture to acquire, share and reuse knowledge. To solve those problems, this paper presented a distributed knowledge intelligence collaborative processing model which uses ontology to build knowledge processing architecture and to integrate domain knowledge distributed on different systems. A medical application model using ontology was presented. It is clear that the model, using ontology to access, reuse, share and organize domain knowledge, can reduce the complexity of distributed knowledge interaction and enhance the merge and cooperation of heterogeneous knowledge sources. [ABSTRACT FROM AUTHOR]