Papers by Author: Hui Yan Jiang

Abstract: Aiming at avoiding misregistration in complicated medical image registration based on SURF (Speed-Up Robust Features)-TPS (Thin-Plate Spline), we propose a novel algorithm. This method is based on SURF and human interaction method for feature extraction. Then we improve SURF-TPS and propose an algorithm named TPS-SEMISURF which obtains the deformation field by calculating the Thin-plate spline of the feature points, and finally does the medical image non-rigid registration according to the parameters. Experimental results showed that the proposed method can register medical images effectively. It has a good robustness and owns better precision and rate than traditional algorithm.

Abstract: This paper proposes a new skin image detection method. First, skin pixel histogram in RGB color space is analyzed. Then Gaussian Mixture Model is used to constructed distribution of skin pixels. Second, a Gaussian parameter combination and selection procedure is implemented with Genetic Algorithms, and the optimal Gaussian Mixture Model can be obtained. Experimental results on public database show that our proposed method outperforms the traditional method with ROC test.

Abstract: In this paper ACO (Ant Colony Optimization) algorithm, which is a well-known intelligent optimization method, is applied to selecting parameters for SVM.ACO has the characteristics of positive feedback, parallel mechanism and distributed computation. This paper gives comparison of ACO-SVM, PSO-SVM whose parameters are determined by particle swarm optimization algorithm, and traditional SVM whose parameters are decided through trial and error. The experimental results on real-world datasets show that this proposed method avoids randomness and subjectivity in the traditional SVM. Additionally it is able to gain better parameters which could dedicate to a higher classification accuracy than the PSO-SVM. Results confirm that proposed optimization method is better than the two others.

Abstract: Snakes are extensively used in computer vision and Image processing. However, when it comes to the liver segmentation from computed tomography (CT) image, the application of the models is limited because it can not extend to certain boundary indentations of the liver. In order to solve this problem, we developed an improved GVF snake model by adding an external force field which can efficiently attract the initial contour to these depression areas, such as the top of the left lobe of liver. The proposed method includes two steps. Firstly, combined with the threshold method and the morphology operation, our model can acquire the initial contour of the liver. Secondly, we create an imposed external force field through the interaction with the system, and we make the initial contour converge under the influence of both GVF field and imposed external force field to get the accurate contour of the liver. The application of this method on abdominal CT image is demonstrated, both qualitatively and quantitatively.

Abstract: A novel edge-based active contour model (ACM) is proposed in this paper. Our edge-based active contour model has many advantages over the conventional active contour models. Firstly, the proposed model can get much smoother contour and needs much less iterations to evolution by being implemented with a special processing named Selectively Binary and Gaussian Filtering Regularized Level Set (SBGFRLS) method. Secondly, we introduce Bilateral Gaussian Filter which can preserve edges to smooth images. So we make weak edges more clear than traditional Gaussian Filter. Thirdly, the level set function can be easily initialized with binary function, which is more efficient to construct than the widely used signed distance function (SDF) because of the special processing. Experiments on synthetic image and segmenting liver from abdominal CT images demonstrate the advantages of the proposed method over geodesic active contours (GAC) in term of both efficiency and accuracy.