Aiming at the problem that the interested target is disappeared or occluded easily in large range of complex scenes by using the traditional camera, a cylinder unwrapping and real-time target tracking method based on panoramic camera is proposed. In this paper, an improved unwrapping algorithm is adopted to transform the panoramic image from omni-directional image. This algorithm effectively solves the distortion problem of panoramic image, and then CamShift combing with Kalman filter algorithm is used to track the moving target. Experimental results demonstrate that the proposed algorithm realizes a real-time and robustness target tracking under large-scale and complex scenes, which contains moving target occluded, temporary disappearance or interference from objects with same color.
In the myocardial perfusion image, the location and shape of the heart change with respiration and heartbeat. Therefore, it is necessary to compensate the movement of positions of myocardial in Cardiac Magnetic Resonance(CMR). To address the poor feature problems in medical image, this paper introduces Markov Random Field(MRF) to tackle this problem and to assess the cardiac movement. According to the neighbor information and intensity information of image pixel blocks in the sequence of cardiac cycle images, the motion vectors can be calculated and the most similar pixel blocks are placed to almost the same position to compensate cardiac movement. Due to complexity of the calculation in MRF, some GPU based methods are introduced to improve computing performance of the whole algorithm. Experimental results demonstrate that the method can effectively correct the movement and deformation of the myocardial perfusion image. The calculation performance increases 400%, the calculation time is one third of the CPU based methods after applying GPU.
Aiming at the problem of the heavy-tailed characteristics in the actual face image, a face recognition method of multi-classification based on mixed Kotz-type distribution is proposed. Mixed Kotz-type distribution and generalized inverse gamma distribution are often used to represent heavy-tailed characteristics. Based on kernel method and probability statistics, this method adjusts the mixed Kotz-type distribution of the parameters to estimate the facial image in the case of heavy-tailed noise tailing. Varying degrees of heavy-tailed noise are added respectively to the ORL face database, Yale face database, Randface(homemade) face database, and a new heavy-tailed noise with varying degrees of face database is formed. Through the verifying of three face database containing different level heavy-tailed noise, results show that the method can estimate the face image trailing feature containing heavy-tailed noise, and has a higher recognition rate.
The fusion algorithm of Mel Frequency Cepstral Coefficient(MFCC) and Linear Prediction Cepstrum Coeficient(LPCC) can only react the static characteristics of the speech and LPCC can not describe the local characteristics of the speech low frequency well. So the fusion of Hilbert-Huang Transform(HHT) cepstrum coefficient and Relative Spectra-Perception Linear Prediction Cepstrum Coefficient(RASTA-PLPCC) is proposed, getting a new speaker recognition algorithm that reflects both vocal mechanism and human ear perceptual characteristics. The HHT cepstrum coefficient reflects the human vocal mechanism, and it can reflect the dynamic characteristics of the speech, as well as better describe the local characteristics of the speech low frequency. PLPCC reflects the human ear perceptual characteristics, whose identification performance is better than the MFCC. Two features are combined with the three fusion algorithms, and the fusion feature is sent into the Gaussian mixture model to do speaker recognition. Simulation results demonstrate that compared with the fusion of LPCC and MFCC, the fusion algorithm gets higher recognition rate, and recognition rate is increased by 8.0%.
Probabilistic model is a valid tool to solve the problem of uncertainty inference and data analysis. An improved algorithm based on Markov network is proposed, which focuses on the uncertainty of ontology matching. The similarity matrix is computed using several traditional algorithms, then the similarity propagation rule is improved, and two structure stability constraint rules and one Disjoint coherence constraint rule are added. The corresponding clique potentials are defined. According to the similarity matrix and these rules, a method to construct the Markov network is proposed. The results of ontology matching are obtained from the posterior probability, which is computed by doing approximate reasoning of the Loopy Belief Propagation(LBP) algorithm. Experimental results on OAEI 2010 show that the algorithm can reduce the complexity of probabilistic model effectively compared with iMatch ontology matching system, meanwhile such various clique rules and the corresponding potentials can increase the precision and the recall rate.