Lixin Zhen

Weightlessness feature — a novel feature for single tri-axial accelerometer based activity recognition

In this paper, a novel weightlessness feature for activity recognition from a tri-axial acceleration signals have been proposed. Since the orientation between accelerometer and userpsilas body may continuously change when user perform activities, we propose an algorithm to calibrate the actual vertical direction of accelerometer signal through estimating the gravitational direction. We combine peaks of signal and weightlessness feature to produce six dimensional weightlessness-based features for activity recognition. Classification of the activities is performed with Support Vector Machine (SVM). The average accuracy of four activities using the proposed weightlessness-based features is 97.21%, which are better than using traditional widely used time-domains features (mean, standard deviation, energy and correlation of acceleration data). Experimental results show that the new features can be used to effectively recognize different human activities and they are robust for different location of accelerometer.

Gesture recognition based on 3D accelerometer for cell phones interaction

This paper proposes a gesture recognition system based on single tri-axis accelerometer mounted on a cell phone for human computer interaction (HCI). Three feature extraction methods, namely discrete cosine transform (DCT), fast Fourier transform (FFT) and a hybrid approach which combine wavelet packet decomposition (WPD) with fast Fourier transform are proposed. Recognition of the gestures is performed with support vector machine (SVM). Recognition results are based on acceleration data collect from 67 subjects. The best average recognition result (87.36%) for 17 complex gestures is achieved with wavelet-based method, while DCT and FFT produce accuracy of 85.16% and 86.92% respectively. The performance of experimental results shows that gesture-based interaction can be used as a novel HCI for mobile applications, such as games control and music navigation.

Design and implementation of a card reader based on build-in camera

With the availability of high-resolution cameras and increased computation power, it becomes possible to implement OCR applications such as business card readers in the mobile device. In this paper, we introduced the design and implementation of a business card reader based on a built-in camera. In order to deal with the challenge of limited resources in a mobile device, we proposed a new method based on multi-resolution analysis of document images. This method improves computation speed and reduces memory requirement of the image-processing step by detecting the text areas in the downscaled image and then analyzing each detected area in the original image. For the OCR engine, we used a two-layer classifier to improve speed. Our experiment gives a satisfactory result.

A NOVEL VISION-BASED FINGER-WRITING CHARACTER RECOGNITION SYSTEM

A novel video-based finger writing virtual character recognition system (FVCRS) is described in this paper. With this FVCR system, one can enter characters into a computer by just using the movement of fingertip, without any additional device such as a keyboard or a digital pen. This provides a new wireless character-inputting method. A simple but effective background model is built for segmenting human-finger movements from cluttered background. A robust fingertip detection algorithm based on feature matching is given, and recognition of the finger-writing character is by a DTW-based classifier. Experiments show that the FVCRS can successfully recognize finger-writing uppercase and lowercase English alphabet with the accuracy of 95.3% and 98.7%, respectively.

Online Chinese character recognition system with handwritten Pinyin input

We have developed a novel online Chinese handwriting recognition system that can recognize a Chinese character either by its handwritten script or by its handwritten Pinyin syllable. The new system is particularly useful when the user forgets how to write the desired character or when the desired character is too complex to be written conveniently. To assure the accuracy and robustness, several classifiers with different characteristics are integrated. The experimental results show that we have achieved an accuracy of 92.5% for 6763-class freely-written Chinese characters and 87.1% for 412-class unconstrained-style Pinyin syllables.

One stroke cursive character recognition using combination of directional and positional features

The paper proposes a new hybrid approach of directional and positional features for on-line one stroke cursive character recognition based on a dynamic time warping (DTW) algorithm. In our camera based user interface, a user inputs various kinds of characters, including Chinese characters, by moving a fingertip. All strokes of the character are connected, so our recognizer is designed for one stroke cursive character recognition. A quadratic curve equation for a local distance measure is employed in DTW to improve the robustness of the classifier, especially for complicated characters. By reconstructing a positional feature from a directional feature, only directional vectors need to be recorded. Thus, the template file size can be reduced a lot. As the template size is small (about 300 K including Chinese characters) and the templates can be easily customized by a user, the recognizer is suitable for hand-held devices. The efficiency of our approach is demonstrated by the promising experimental results.

Camera based mixed-lingual card reader for mobile device

With the availability of high-resolution cameras and increased computation power, it becomes possible to implement OCR applications such as business card reader in the mobile device. In this paper we introduced the design and implementation of a mixed-lingual business card reader based on built-in camera. It has the capability to recognize business cards with Chinese or English characters. In order to deal with the challenge of limited resource in mobile device, we proposed some new methods to reduce the resource requirement of the image processing and the Chinese OCR engine. Our experiment gives satisfactory result.

Kernel modified quadratic discriminant function for facial expression recognition

The Modified Quadratic Discriminant Function was first proposed by Kimura et al to improve the performance of Quadratic Discriminant Function, which can be seen as a dot-product method by eigen-decompostion of the covariance matrix of each class. Therefore, it is possible to expand MQDF to high dimension space by kernel trick. This paper presents a new kernel-based method to pattern recognition, Kernel Modified Quadratic Discriminant Function(KMQDF), based on MQDF and kernel method. The proposed KMQDF is applied in facial expression recognition. JAFFE face database and the AR face database are used to test this algorithm. Experimental results show that the proposed KMQDF with appropriated parameters can outperform 1-NN, QDF, MQDF classifier.

An efficient candidate set size reduction method for coarse-classification in Chinese handwriting recognition

In this paper, we introduce an efficient clustering based coarse-classifier for a Chinese handwriting recognition system to accelerate the recognition procedure. We define a candidate-cluster-number for each character. The defined number indicates the within-class diversity of a character in the feature space. Based on the candidate-cluster-number of each character, we use a candidate-refining module to reduce the size of the candidate set of the coarse-classifier. Experiments show that the method effectively reduces the output set size of the coarse-classifier, while keeping the same coverage probability of the candidate set. The method has a low computation-complexity.

A Probabilistic Framework for Soft Target Learning in Online Cursive Handwriting Recognition

To develop effective learning algorithms for online cursive word recognition is still a challenge research issue. In this paper, we propose a probabilistic framework to model the inherent ambiguity of cursive handwriting by using soft target vector of each character class. In the proposed algorithm, the values of soft targets are estimated by introducing a lower bound on the log likelihood and optimizing this lower bound via an EM like algorithm. In the experiments on 207K collected cursive words written by 1060 subjects, the proposed algorithm clearly outperforms baseline method with word error reduction up to 11.6%. Furthermore, the estimated soft target values are useful for measuring the separability between output classes.