Chenhui Yang

Simple Vehicle Detection with Shadow Removal at Intersection

Moving vehicle detection is an important part in intelligent transportation system applications. The purpose is to track each moving vehicle in the video frames. This paper analyzes the background generation and shadow removal methods in traditional background subtraction approach, and presents a simple algorithm to generate and update color background image based on statistical method at intersection. Our approach of shadow removal is based on both NCC (Normalized Cross-correlation) and frame difference. Experiments showed that this detection method can well detect vehicles at heavy traffic intersection.

Symmetry Detection for Multi-object Using Local Polar Coordinate

In this paper, a novel method is presented, which detects symmetry axes for multi-object. It uses vertical line detection method in local polar coordinate. The approach costs little computation and can get efficient results, which means that it can be used in database applications.

Marine Phytoplankton Recognition Using Hybrid Classification Methods

Marine phytoplanktons are unicellular algae with a variety of shapes and ornamentation, and they are widely used as indicators of marine ecosystem changes. A dual layer and hybrid classifier is presented in this study for phytoplankton recognition. The method is based on k-NN, SVM mechanisms and uses shape and texture information such as moments, geometric features and gray level co-occurrence matrix features. Each individual classifier has its own specific input feature and decision mechanism. The marine phytoplankton recognition experiment shows that the proposed classification method outperforms two well-known stand-alone classifiers, k-NN and SVM.

Automatic Identification of Round Diatom

In this work, a method for automatic identification of round diatoms based on image texture features is presented. This method combines segmentation adjustment by curve fitting and texture measures based on the spectrum features. The classification accuracy was tested using leave on out methods. With classification carried out using a BP neural network we attained 96.2% accuracy from a set of image containing six species of round diatom. The result is an effective attempt of round diatom identification based on texture character.

A Novel Symmetry Detection Method for Images and Its Application for Motion Deblurring

In this paper, we provide a novel method for symmetry detection in diatom images based on spectrum analysis. We notice that the spectrum of an image has naturally zero point symmetry and that the direction of symmetry axis in an image is the same with that in the spectrum of the image. Our method detects perfect and imperfect with more flexibility and efficiency than the traditional ones. Using Fourier transformation, we turn the imperfect symmetry image into perfect symmetry image, in which the symmetry can be easily detected. And we use some algorithms to locate the symmetry axes, which can save much time and computation. And we give a model application of our method in the field of motion deblurring.

A Svm-Based Algorithm for Automatic Species Classification of a Marine Diatom Genus Coscinodiscus Ehrenberg

Coscinodiscus Ehrenberg is a large and ecologically important diatom genus with plentiful species in marine phytoplankton and with a variety of round shapes and ornamentation. These properties can be measured by computer image pre-segmentation and feature extraction with threshold methods. However，it proves to be complicated task because of the high spatial variability of ornamentation properties. Researchers have shown a Teach-program and a number of library functions operating on sample image lists (SILs) and operating on classifiers (CFs) to solve the problem. In this paper, we present Coscinodiscus Ehrenberg ornamentation classifier algorithm called support vector machines (SVMs) to derive a new set of SILs and CFs. The principal purpose of SVMs is Coscinodiscus Ehrenberg images pattern recognition approach. A pattern is in this context always the SILs contained in a sub-rectangle of some given (possibly larger) image. For the same classifier this sub-rectangle must always have the same dimensions, while the query image to be searched may be arbitrarily large. The training is done by preparing SILs for the pattern taxa in question and feeding them to CFs created. Our classifier generation with preprocessing code optimization achieves a AAAAA preprocessing code, a 0.981 learning success, a 100% computational complexity. Train with SILs achieves 212 samples, 17 taxa, a 0.472% error rate and Test with Query image searching achieves 253 samples，17 taxa，a 15.81% error rate. The experiments demonstrate that the proposed method is very robust to the threshold segmentation and ornamentation feature extraction of Coscinodiscus Ehrenberg images, and is effective and useful for species classification of Coscinodiscus Ehrenberg.

Improving Contour Tracker through Evolutionary Optimization

Tracking contours in an image sequence is a challenging task. Tracking algorithms based on particle filter have been proposed for this nonlinear problem. But, contour trackers often collapse due to the sample impoverishment of the traditional particle filter. In this paper, we integrate evolutionary optimization into particle filter, and it is applied to visual contour tracking. The impoverishment problem can be prevented using crossover and mutation operation. Moreover, the re-sampling process is replaced by selection operation. Particles can be redistributed to the local modes with the evolution of the particle population. Experimental results on some recorded videos demonstrate the proposed tracker has the better performance for the changed contour and the clutter.