Mei Xie

Preceding vehicle detection using Histograms of Oriented Gradients

This paper presents a monocular vision-based preceding vehicle detection system using Histogram of Oriented Gradient (HOG) based method and linear SVM classification. Our detection algorithm consists of three main components: HOG feature extraction, linear SVM classifier training and vehicles detection. Integral Image method is adopted to improve the HOG computational efficiency, and hard examples are generated to reduce false positives in the training phase. In detection step, the multiple overlapping detections due to multi-scale window searching are very well fused by non-maximum suppression based on mean-shift. The monocular system is tested under different traffic scenarios (e.g., simply structured highway, complex urban environments, local occlusion conditions), illustrating good performance.

Iris Feature Extraction Based on Wavelet Packet Analysis

Feature extraction is the crucial part of an iris recognition system. Multi-resolution analysis such as Gabor filters and wavelet transform (WT) help to characterize different scales of iris texture. In this paper, we propose a novel multi-resolution approach based on wavelet packets transform (WPT) for iris texture analysis and recognition. The development of this approach is motivated by the observation that dominant frequencies of iris texture are located in the low and middle frequency channels. With an adaptive threshold, we quantize WPT subimages coefficients into 1, 0 or -1 as iris signature. This signature presents the local information of different irises. Manhattan Distance is used to measure the degree of dissimilarity between two iris sets. Experimental results show that the algorithm is efficient to describe local information

A Modified Skull-Stripping Method Based on Morphological Processing

In the paper, we first discussed the method of skull-stripping based on edge detection in detail. The method can extract the brain tissue from a normal MRI (magnetic resonance image) quickly and accurately. But if there are tumors at the boundary of the brain tissues whose intensity is different with normal tissues’, some unwanted edge will appear while detecting the boundary of the brain tissue. So, we proosed a modified method to solve this issue in this paper. Our procedure consists of three steps. First,the MRI is processed with an anisotropic diffusion filter to encourage intraregion smoothing in preference to smoothing across the boundaries. And then find the region between the skull and the brain tissue using thresholding methods. It is difficult to remove the unwanted edges directly,but we found that when we processed it using a skeletonization algorithm, the unwanted edges became the branches on the skeleton of the region. Finally, use a sequence of morphological and connected component operator to make sure the region is closed. The closed region is that we wish for. Results on 2-D MR images are provided.

A Central Sub-image Based Global Motion Estimation Method for In-Car Video Stabilization

This paper presents a novel global motion estimation method based on the phase correlation of central sub-image. In this study, we consider the case that the In-Car videos are captured from the cameras placed in front of a car. The backgrounds of these In-Car videos usually vary with the moving car, which result in the inaccuracy of classical image stabilization method. As a result, a central sub-image based image stabilization method is presented in this paper. The simulation results show that the proposed method is efficient in improving the accuracy of detecting global motion vectors of In-Car videos.

Fingerprint segmentation based on PCNN and morphology

As an important step in an automatic fingerprint recognition system, fingerprint segmentation aims to extract the foreground of a fingerprint image in an efficient way. In this paper, an initiative algorithm for fingerprint segmentation is presented. First, the model of Pulse Coupled Neural Networks (PCNN) is utilized to binarize the fingerprint image. Then, morphological methods are adopted to obtain compact clusters of the binary fingerprint image. Since there might be other interfering regions after morphological operation, we also use the labeling method to find the largest compact cluster as the foreground region of the fingerprint image. Experimental results show that this method is robust to the complicated backgrounds of fingerprint images while keeping a smooth contour of the foreground region.

Shear-Warp Rendering Algorithm Based on Radial Basis Functions Interpolation

The shear-warp algorithm is considered as one of the most fastest algorithms in volume rendering. However, it achieves such speed at the price of reduced image quality. During the shear-warp resampling process, the sampling distance varies as the viewing orientation varies, and at some viewing angles, the more larger distance would result in the image-aliasing phenomenon in the resulting image and affect the rendering quality of image. To weaken the aliasing phenomenon effectly, this paper present a new interpolation method based on multiquadric method which is one of the radial basis functions interpolation. In this method, we construct a curve from four original and adjacent sampling points along the prime view orientation, and then a new sampling point is obtained in this curve and used to construct new slices, which are added into original slices. Finally, a new 3D data field can be got. Based on this new data field, shear-warp rendering could significantly improve the image quality and obviously reduce the aliasing phenomenon.

Lane detection using Heuristic Search Methods based on color clustering

Concerning the problem of lane detection in the Lane Departure Warning (LDW) system, this paper uses Heuristic Search Methods to detect the lane boundary points based on the CIELab color features clustering. Color space providing us more precious information than gray scale, and Heuristic Search Methods narrowing the search region, this algorithm is robust and adaptive to the actual road conditions. According to the geometry structure of super highway, quadratic curve is adopted to match the lane. At last, use Simulink, the Signal Processing Blockset, and the Video and Image Processing Blockset developing lane detection and tracking model.

An adaptive digital image stabilization based on empirical mode decomposition

An adaptive digital image stabilization method based on empirical mode decomposition is presented in this paper. The frame position signal is firstly decomposed into a series of intrinsic mode functions, and then the intrinsic mode functions that represent the intentional camera motions are adaptively selected to reconstruct the smoothed frame position signal. The experiments show that the proposed digital image stabilization method not only successfully removes the unwanted camera motions but also is parameter independent.

Chinese character recognition based on character reconstruction

A completely new and effective algorithm for Chinese character recognition is proposed in this paper. The recognition is based on character reconstruction. And then we can obtain a new normalized character. Through character reconstruction we can reduce the error brought by the blur and the tilt of the original image. We first obtain the structure information of the original character and then we reconstruct a new character according to our rules. The structure information includes the information of the horizontal lines, the vertical line, the bias lines and dot. Then we compare the sample character and the template character and obtain the recognition result. A large number of experiments prove the robustness and high performance of the algorithm. The recognition rate of this algorithm is 96%. It can bear images blur and tilt.

Face location with LBP scale transform

Local Binary Patterns (LBP) is an effective texture description operator and the histogram that it generates has been proved to be a very useful texture feature to adapt to rotation and illumination. Using the LBP features as feature vectors in adaBoost classifier for target identification has become a trend. But LBP is bound by the scale transformation, so it is not widely used in adaBoost face detector. This paper proposes a scale transform formula for Local Binary Patterns. Based on this formula, LBP features extracted from single fixed size templates can be trained to identify any size of faces. This paper also proposes a method to obtain particular detecting sub-areas called binary ring-shaped sub-windows, which can keep the LBP features rotation invariant. Experimental results show that the method we proposed here is feasible in face detecting.